TupleArray::TupleArray(ArrayDesc const& schema, vector< boost::shared_ptr<Tuple> > const& data, Coordinate offset)
: desc(schema),
start(schema.getDimensions()[0].getStart() + offset),
end(start + offset + schema.getDimensions()[0].getLength() - 1),
tuples(data), chunkSize(schema.getDimensions()[0].getChunkInterval())
{
desc.cutOverlap();
if (Coordinate(start + tuples.size()) <= end) {
end = start + tuples.size() - 1;
}
}
void fillUsedPlugins(const ArrayDesc& desc, vector<string>& plugins) const
{
for (size_t i = 0; i < desc.getAttributes().size(); i++) {
const string& libName = TypeLibrary::getTypeLibraries().getObjectLibrary(desc.getAttributes()[i].getType());
if (libName != "scidb")
plugins.push_back(libName);
}
for (size_t i = 0; i < desc.getDimensions().size(); i++) {
const string& libName = TypeLibrary::getTypeLibraries().getObjectLibrary(desc.getDimensions()[i].getType());
if (libName != "scidb")
plugins.push_back(libName);
}
}
SplitArray::SplitArray(ArrayDesc const& desc,
const boost::shared_array<char>& src,
Coordinates const& from,
Coordinates const& till,
shared_ptr<Query>const& query)
: DelegateArray(desc, shared_ptr<Array>(), true),
_startingChunk(from),
_from(from),
_till(till),
_size(from.size()),
_src(src),
_empty(false)
{
assert(query);
_query = query;
desc.getChunkPositionFor(_startingChunk);
Dimensions const& dims = desc.getDimensions();
for (size_t i = 0, n = dims.size(); i < n; i++) {
_size[i] = _till[i] - _from[i] + 1;
if (_size[i] == 0) {
_empty = true;
}
if (_till[i] > dims[i].getEndMax()) {
_till[i] = dims[i].getEndMax();
}
}
}
TupleArray::TupleArray(ArrayDesc const& schema, vector< boost::shared_ptr<ConstArrayIterator> > const& arrayIterators, size_t shift, size_t step)
: desc(schema),
start(schema.getDimensions()[0].getStart()),
end(schema.getDimensions()[0].getEndMax()),
chunkSize(schema.getDimensions()[0].getChunkInterval())
{
if (schema.getDimensions().size() != 1)
throw USER_EXCEPTION(SCIDB_SE_EXECUTION, SCIDB_LE_MULTIDIMENSIONAL_ARRAY_NOT_ALLOWED);
append(arrayIterators, shift, step);
if (start == MIN_COORDINATE || end == MAX_COORDINATE) {
start = 0;
end = tuples.size()-1;
} else if (Coordinate(start + tuples.size()) <= end) {
end = start + tuples.size() - 1;
}
}
ArrayDesc InputArray::generateShadowArraySchema(ArrayDesc const& targetArray, string const& shadowArrayName)
{
Attributes const& srcAttrs = targetArray.getAttributes(true);
size_t nAttrs = srcAttrs.size();
Attributes dstAttrs(nAttrs+2);
for (size_t i = 0; i < nAttrs; i++) {
dstAttrs[i] = AttributeDesc(i, srcAttrs[i].getName(), TID_STRING, AttributeDesc::IS_NULLABLE, 0);
}
dstAttrs[nAttrs] = AttributeDesc(nAttrs, "row_offset", TID_INT64, 0, 0);
dstAttrs[nAttrs+1] = AttributeDesc(nAttrs+1, DEFAULT_EMPTY_TAG_ATTRIBUTE_NAME,
TID_INDICATOR, AttributeDesc::IS_EMPTY_INDICATOR, 0);
return ArrayDesc(shadowArrayName, dstAttrs, targetArray.getDimensions());
}
//param desc --> the input array schema
ArrayDesc createWindowDesc(ArrayDesc const& desc)
{
//get dimensions for output array
Attributes const &attrs = desc.getAttributes();
/*
Dimensions aggrDims(dims.size());
for (size_t i = 0; i < dims.size(); i++)
{
DimensionDesc const& srcDim = dims[i];
aggrDims[i] = DimensionDesc(srcDim.getBaseName(),
srcDim.getNamesAndAliases(),
srcDim.getStartMin(),
srcDim.getCurrStart(),
srcDim.getCurrEnd(),
srcDim.getEndMax(),
srcDim.getChunkInterval(),
0);
}
*/
Attributes newAttributes;
size_t n = 0;
for (size_t i=desc.getDimensions().size()*2; i < _parameters.size()-1; i=i+2)
{
const AttributeDesc &attr = attrs[((boost::shared_ptr<OperatorParamReference>&)_parameters[i])->getObjectNo()];
newAttributes.push_back(AttributeDesc(n, attr.getName(),
attr.getType(),
attr.getFlags(),
attr.getDefaultCompressionMethod(),
attr.getAliases()));
}
return ArrayDesc(desc.getName(), newAttributes, desc.getDimensions());
}
//param desc --> the input array schema
inline ArrayDesc createWindowDesc(ArrayDesc const& desc)
{
//get dimensions for output array
Dimensions const& dims = desc.getDimensions();
Dimensions aggrDims(dims.size());
for (size_t i = 0; i < dims.size(); i++)
{
DimensionDesc const& srcDim = dims[i];
aggrDims[i] = DimensionDesc(srcDim.getBaseName(),
srcDim.getNamesAndAliases(),
srcDim.getStartMin(),
srcDim.getCurrStart(),
srcDim.getCurrEnd(),
srcDim.getEndMax(),
srcDim.getChunkInterval(),
0);
}
ArrayDesc output(desc.getName(), Attributes(), aggrDims);
//get the aggregates, check if they make sense, make attributes for output array
//_parameters[0~dims.size()*2-1] --> window boundaries, already get in inferSchema
for (size_t i = dims.size()*2; i < _parameters.size(); i++)
{
boost::shared_ptr<scidb::OperatorParam> param = _parameters[i];
if ( param->getParamType() != PARAM_AGGREGATE_CALL) {
throw USER_QUERY_EXCEPTION(SCIDB_SE_INFER_SCHEMA,
SCIDB_LE_OP_WINDOW_ERROR5,
_parameters[i]->getParsingContext());
}
addAggregatedAttribute( (shared_ptr<OperatorParamAggregateCall> &) param, desc, output, true);
}
if ( desc.getEmptyBitmapAttribute()) //?
{
AttributeDesc const* eAttr = desc.getEmptyBitmapAttribute();
output.addAttribute(AttributeDesc(output.getAttributes().size(),
eAttr->getName(),
eAttr->getType(),
eAttr->getFlags(),
eAttr->getDefaultCompressionMethod()));
}
return output;
}
FITSInputArray::FITSInputArray(ArrayDesc const& array, string const& filePath, uint32_t hdu, std::shared_ptr<Query>& query)
: parser(filePath),
hdu(hdu),
desc(array),
dims(array.getDimensions()),
nDims(dims.size()),
nAttrs(array.getAttributes(true).size()),
values(nAttrs),
chunks(nAttrs),
chunkIterators(nAttrs),
chunkIndex(0),
chunkPos(nDims),
query(query)
{
initValueHolders();
// Most initialization steps are only done later, when the first
// chunk is requested by an iterator. See getChunkByIndex()
}
inline ArrayDesc createWindowDesc(ArrayDesc const& desc)
{
Dimensions const& dims = desc.getDimensions();
Dimensions aggDims(dims.size());
for (size_t i = 0, n = dims.size(); i < n; i++)
{
DimensionDesc const& srcDim = dims[i];
aggDims[i] = DimensionDesc(srcDim.getBaseName(),
srcDim.getNamesAndAliases(),
srcDim.getStartMin(),
srcDim.getCurrStart(),
srcDim.getCurrEnd(),
srcDim.getEndMax(),
srcDim.getChunkInterval(),
0,
srcDim.getType(),
srcDim.getFlags(),
srcDim.getMappingArrayName(),
srcDim.getComment(),
srcDim.getFuncMapOffset(),
srcDim.getFuncMapScale());
}
ArrayDesc output (desc.getName(), Attributes(), aggDims);
for (size_t i = dims.size() * 2, size = _parameters.size(); i < size; i++)
{
addAggregatedAttribute( (shared_ptr <OperatorParamAggregateCall> &) _parameters[i], desc, output);
}
if ( desc.getEmptyBitmapAttribute())
{
AttributeDesc const* eAtt = desc.getEmptyBitmapAttribute();
output.addAttribute(AttributeDesc(output.getAttributes().size(), eAtt->getName(),
eAtt->getType(), eAtt->getFlags(), eAtt->getDefaultCompressionMethod()));
}
return output;
}
ArrayDesc inferSchema(std::vector< ArrayDesc> schemas, boost::shared_ptr< Query> query)
{
assert(schemas.size() == 1);
ArrayDesc const& srcDesc = schemas[0];
ArrayDesc dstDesc = ((boost::shared_ptr<OperatorParamSchema>&)_parameters[0])->getSchema();
//Compile a desc of all possible attributes (aggregate calls first) and source dimensions
ArrayDesc aggregationDesc (srcDesc.getName(), Attributes(), srcDesc.getDimensions());
vector<string> aggregatedNames;
//add aggregate calls first
for (size_t i = 1; i < _parameters.size(); i++)
{
addAggregatedAttribute( (shared_ptr <OperatorParamAggregateCall>&) _parameters[i], srcDesc, aggregationDesc);
aggregatedNames.push_back(aggregationDesc.getAttributes()[aggregationDesc.getAttributes().size()-1].getName());
}
//add other attributes
BOOST_FOREACH(const AttributeDesc &srcAttr, srcDesc.getAttributes())
{
//if there's an attribute with same name as an aggregate call - skip the attribute
bool found = false;
BOOST_FOREACH(const AttributeDesc &aggAttr, aggregationDesc.getAttributes())
{
if( aggAttr.getName() == srcAttr.getName())
{
found = true;
}
}
if (!found)
{
aggregationDesc.addAttribute(AttributeDesc( aggregationDesc.getAttributes().size(),
srcAttr.getName(),
srcAttr.getType(),
srcAttr.getFlags(),
srcAttr.getDefaultCompressionMethod(),
srcAttr.getAliases(),
&srcAttr.getDefaultValue(),
srcAttr.getDefaultValueExpr(),
srcAttr.getComment(),
srcAttr.getVarSize()));
}
}
//Ensure attributes names uniqueness.
if (!dstDesc.getEmptyBitmapAttribute())
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_OP_REDIMENSION_ERROR1);
BOOST_FOREACH(const AttributeDesc &dstAttr, dstDesc.getAttributes())
{
BOOST_FOREACH(const AttributeDesc &srcAttr, aggregationDesc.getAttributes())
{
if (srcAttr.getName() == dstAttr.getName())
{
if (srcAttr.getType() != dstAttr.getType())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_ATTRIBUTE_TYPE)
<< srcAttr.getName() << srcAttr.getType() << dstAttr.getType();
}
if (!dstAttr.isNullable() && srcAttr.isNullable())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_ATTRIBUTE_FLAGS)
<< srcAttr.getName();
}
goto NextAttr;
}
}
BOOST_FOREACH(const DimensionDesc &srcDim, aggregationDesc.getDimensions())
{
if (srcDim.hasNameOrAlias(dstAttr.getName()))
{
if (dstAttr.getType() != TID_INT64)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_DESTINATION_ATTRIBUTE_TYPE)
<< dstAttr.getName() << TID_INT64;
}
if (srcDim.getType() != TID_INT64)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_SOURCE_DIMENSION_TYPE)
<< dstAttr.getName() << TID_INT64;
}
if (dstAttr.getFlags() != 0)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_DESTINATION_ATTRIBUTE_FLAGS)
<< dstAttr.getName();
}
goto NextAttr;
}
}
if (dstAttr.isEmptyIndicator() == false)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_UNEXPECTED_DESTINATION_ATTRIBUTE)
<< dstAttr.getName();
}
NextAttr:;
}
BOOST_FOREACH(const DimensionDesc &dstDim, dstDesc.getDimensions())
{
if (dstDim.getChunkOverlap() != 0)
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_OP_REDIMENSION_STORE_ERROR3);
BOOST_FOREACH(const AttributeDesc &srcAttr, aggregationDesc.getAttributes())
{
if (dstDim.hasNameOrAlias(srcAttr.getName()))
{
for (size_t i = 0; i< aggregatedNames.size(); i++)
{
if (srcAttr.getName() == aggregatedNames[i])
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_OP_REDIMENSION_ERROR2);
}
if (srcAttr.getType() != TID_INT64)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_SOURCE_ATTRIBUTE_TYPE)
<< srcAttr.getName() << TID_INT64;
}
if (srcAttr.getFlags() != 0)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_SOURCE_ATTRIBUTE_FLAGS)
<< srcAttr.getName() << TID_INT64;
}
if (dstDim.getType() != TID_INT64)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_DESTINATION_DIMENSION_TYPE)
<< srcAttr.getName() << TID_INT64;
}
goto NextDim;
}
}
BOOST_FOREACH(const DimensionDesc &srcDim, aggregationDesc.getDimensions())
{
if (srcDim.hasNameOrAlias(dstDim.getBaseName()))
{
if (dstDim.getType() != srcDim.getType() ||
dstDim.getStart() != srcDim.getStart() ||
dstDim.getLength() != srcDim.getLength() ||
dstDim.getChunkInterval() != srcDim.getChunkInterval())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_DIMENSIONS_DONT_MATCH)
<< srcDim.getBaseName() << dstDim.getBaseName();
}
goto NextDim;
}
}
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_UNEXPECTED_DESTINATION_DIMENSION) << dstDim.getBaseName();
NextDim:;
}
return ArrayDesc(srcDesc.getName()+"_redimension", dstDesc.getAttributes(), dstDesc.getDimensions(), dstDesc.getFlags());
}
inline size_t nCol(const ArrayDesc& desc, bool transpose=false) { return nCol(desc.getDimensions(), transpose); }
ArrayDesc inferSchema(std::vector< ArrayDesc> schemas, boost::shared_ptr< Query> query)
{
assert(schemas.size() == 1);
ArrayDesc const& srcDesc = schemas[0];
ArrayDesc dstDesc = ((boost::shared_ptr<OperatorParamSchema>&)_parameters[0])->getSchema();
//Compile a desc of all possible attributes (aggregate calls first) and source dimensions
ArrayDesc aggregationDesc (srcDesc.getName(), Attributes(), srcDesc.getDimensions());
vector<string> aggregatedNames;
//add aggregate calls first
for (size_t i = 1; i < _parameters.size(); i++)
{
addAggregatedAttribute( (shared_ptr <OperatorParamAggregateCall>&) _parameters[i], srcDesc, aggregationDesc);
string aggName = aggregationDesc.getAttributes()[aggregationDesc.getAttributes().size()-1].getName();
bool aggFound = false;
BOOST_FOREACH(const AttributeDesc &dstAttr, dstDesc.getAttributes()) {
if (dstAttr.getName() == aggName) {
aggFound = true;
break;
}
}
if (!aggFound) {
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_ATTRIBUTE_DOESNT_EXIST) << aggName << dstDesc.getName();
}
aggregatedNames.push_back(aggName);
}
//add other attributes
BOOST_FOREACH(const AttributeDesc &srcAttr, srcDesc.getAttributes())
{
//if there's an attribute with same name as an aggregate call - skip the attribute
bool found = false;
BOOST_FOREACH(const AttributeDesc &aggAttr, aggregationDesc.getAttributes())
{
if( aggAttr.getName() == srcAttr.getName())
{
found = true;
}
}
if (!found)
{
aggregationDesc.addAttribute(AttributeDesc( aggregationDesc.getAttributes().size(),
srcAttr.getName(),
srcAttr.getType(),
srcAttr.getFlags(),
srcAttr.getDefaultCompressionMethod(),
srcAttr.getAliases(),
&srcAttr.getDefaultValue(),
srcAttr.getDefaultValueExpr(),
srcAttr.getVarSize()));
}
}
//Ensure attributes names uniqueness.
if (!dstDesc.getEmptyBitmapAttribute())
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_OP_REDIMENSION_ERROR1);
BOOST_FOREACH(const AttributeDesc &dstAttr, dstDesc.getAttributes())
{
BOOST_FOREACH(const AttributeDesc &srcAttr, aggregationDesc.getAttributes())
{
if (srcAttr.getName() == dstAttr.getName())
{
if (srcAttr.getType() != dstAttr.getType())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_ATTRIBUTE_TYPE)
<< srcAttr.getName() << srcAttr.getType() << dstAttr.getType();
}
if (!dstAttr.isNullable() && srcAttr.isNullable())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_ATTRIBUTE_FLAGS)
<< srcAttr.getName();
}
goto NextAttr;
}
}
BOOST_FOREACH(const DimensionDesc &srcDim, aggregationDesc.getDimensions())
{
if (srcDim.hasNameAndAlias(dstAttr.getName()))
{
if (dstAttr.getType() != TID_INT64)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_DESTINATION_ATTRIBUTE_TYPE)
<< dstAttr.getName() << TID_INT64;
}
if (dstAttr.getFlags() != 0)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_DESTINATION_ATTRIBUTE_FLAGS)
<< dstAttr.getName();
}
goto NextAttr;
}
}
if (dstAttr.isEmptyIndicator() == false)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_UNEXPECTED_DESTINATION_ATTRIBUTE)
<< dstAttr.getName();
}
NextAttr:;
}
Dimensions outputDims;
size_t nNewDims = 0;
BOOST_FOREACH(const DimensionDesc &dstDim, dstDesc.getDimensions())
{
if (dstDim.getChunkOverlap() > dstDim.getChunkInterval())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_OVERLAP_CANT_BE_LARGER_CHUNK);
}
BOOST_FOREACH(const AttributeDesc &srcAttr, aggregationDesc.getAttributes())
{
if (dstDim.hasNameAndAlias(srcAttr.getName()))
{
for (size_t i = 0; i< aggregatedNames.size(); i++)
{
if (srcAttr.getName() == aggregatedNames[i])
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_OP_REDIMENSION_ERROR2);
}
if ( !IS_INTEGRAL(srcAttr.getType()) || srcAttr.getType() == TID_UINT64 )
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_SOURCE_ATTRIBUTE_TYPE)
<< srcAttr.getName() << TID_INT64;
}
outputDims.push_back(dstDim);
goto NextDim;
}
}
BOOST_FOREACH(const DimensionDesc &srcDim, aggregationDesc.getDimensions())
{
if (srcDim.hasNameAndAlias(dstDim.getBaseName()))
{
DimensionDesc outputDim = dstDim;
outputDims.push_back(outputDim);
goto NextDim;
}
}
//one synthetic dimension allowed
if (nNewDims++ != 0 || !aggregatedNames.empty() )
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_UNEXPECTED_DESTINATION_DIMENSION) << dstDim.getBaseName();
}
outputDims.push_back(dstDim);
NextDim:;
}
return ArrayDesc(srcDesc.getName(), dstDesc.getAttributes(), outputDims, dstDesc.getFlags());
}
shared_ptr< Array > execute(vector< shared_ptr< Array> >& inputArrays, shared_ptr<Query> query)
{
// I maintain the log of the operator in a local file named after Correlation_N.log, N is the instance ID.
stringstream logFileName;
logFileName << "/home/scidb/preselect_" << query->getInstanceID() << ".log";
FILE *logFile;
logFile = fopen(logFileName.str().c_str(), "w");
shared_ptr<Array> originalArray = inputArrays[0];
shared_ptr<Array> correlationArray = inputArrays[1];
ArrayDesc originalSchema = originalArray->getArrayDesc();
ArrayDesc corrSchema = correlationArray->getArrayDesc();
Dimensions originalDims = originalSchema.getDimensions();
Dimensions corrDims = corrSchema.getDimensions();
DimensionDesc originalDimsP = originalDims[1];
DimensionDesc corrDimsP = corrDims[0];
// Note the correlation array doesn't have Y column.
Coordinate p = corrDimsP.getCurrLength();
fprintf(logFile, "p = %ld\n # of chunk = %ld\n", p, corrSchema.getNumberOfChunks());
fflush(logFile);
shared_ptr<ConstArrayIterator> corrArrayIter = correlationArray->getIterator(0);
if(! corrArrayIter->end() )
{
correlation *corr = new correlation[p];
// The correlation array will always have only 1 chunk (we designed correlation array like this), so no loops here.
shared_ptr<ConstChunkIterator> corrChunkIter = corrArrayIter->getChunk().getConstIterator();
for(Coordinate i=0; i<p; ++i)
{
corr[i].id = i+1;
corr[i].corr = corrChunkIter->getItem().getDouble();
//fprintf(logFile, "%d, %f\n", corr[i].id, corr[i].corr);
++(*corrChunkIter);
}
//fflush(logFile);
qsort(corr, p, sizeof(correlation), &comp);
for(Coordinate i=0; i<p; ++i)
{
fprintf(logFile, "%d, %f\n", corr[i].id, corr[i].corr);
}
fflush(logFile);
Coordinate d = ((boost::shared_ptr<OperatorParamPhysicalExpression>&)_parameters[0])->getExpression()->evaluate().getInt64();
fprintf(logFile, "d=%ld\n", d);
stringstream ss;
vector<string> names;
names.push_back("j");
vector<TypeId> types;
types.push_back(TID_INT64);
for(Coordinate i=0; i<d; ++i)
{
ss << "j=" << corr[i].id << " or ";
}
ss << "j=" << p+1;
fprintf(logFile, "%s\n", ss.str().c_str());
fflush(logFile);
Expression e;
e.compile(ss.str(), names, types);
fclose(logFile);
boost::shared_ptr<scidb::Query> emptyQuery;
return boost::shared_ptr<Array>(new FilterArray(_schema, inputArrays[0], boost::make_shared<Expression>(e), emptyQuery, _tileMode));
}
else
{
shared_ptr<Array> outputArray(new MemArray(_schema, query));
fclose(logFile);
return outputArray;
}
}
inline unsigned int chunkCol(const ArrayDesc& desc, bool transpose=false) { return chunkCol(desc.getDimensions(), transpose); }
ReshapeArray::ReshapeArray(ArrayDesc const& desc, boost::shared_ptr<Array> const& array)
: DelegateArray(desc, array),
inDims(array->getArrayDesc().getDimensions()),
outDims(desc.getDimensions())
{
}
shared_ptr< Array > execute(vector< shared_ptr< Array> >& inputArrays, shared_ptr<Query> query) {
shared_ptr<Array> outputArray(new MemArray(_schema, query));
shared_ptr<Array> inputArray = inputArrays[0];
ArrayDesc inputSchema = inputArray->getArrayDesc();
// Get descriptor of two dimensions d and n.
DimensionDesc dimsN = inputSchema.getDimensions()[0];
DimensionDesc dimsD = inputSchema.getDimensions()[1];
size_t n = dimsN.getCurrEnd() - dimsN.getCurrStart() + 1;
// Note: the input data set should have d+1 dimensions (including Y)
size_t d = dimsD.getCurrEnd() - dimsD.getCurrStart();
size_t nStart = dimsN.getCurrStart();
size_t dStart = dimsD.getCurrStart();
// Get chunk size of n.
size_t nChunkSize = dimsN.getChunkInterval();
// Helps to accumulate the n and L.
z_i[0] = 1.0;
shared_ptr<ConstArrayIterator> inputArrayIter = inputArray->getConstIterator(0);
Coordinates chunkPosition;
size_t i, j, k, m;
while(! inputArrayIter->end() ) {
shared_ptr<ConstChunkIterator> chunkIter = inputArrayIter->getChunk().getConstIterator();
chunkPosition = inputArrayIter->getPosition();
for(i=chunkPosition[0]; i<chunkPosition[0] + nChunkSize; i++) {
// In case the chunk is partially filled.
if(i == n + nStart) {
break;
}
for(j=chunkPosition[1], m=1; j<=chunkPosition[1]+d; j++, m++) {
// In case the chunk is partially filled.
if(j == d + 1 + dStart) {
break;
}
z_i[m] = chunkIter->getItem().getDouble();
++(*chunkIter);
}
for(k=0; k<=d+1; ++k) {
// This operator is not optimized for entries with value zero.
// TODO: should use fabs(z_i[k]) < 10e-6
// if(z_i[k] == 0.0) {
// continue;
// }
for(m=0; m<=k; ++m) {
Gamma[k][m] += z_i[k]*z_i[m];
}
}
}
++(*inputArrayIter);
}
/**
* The "logical" instance ID of the instance responsible for coordination of query.
* COORDINATOR_INSTANCE if instance execute this query itself.
*/
if(query->getInstancesCount() > 1) {
if(query->getInstanceID() != 0) {
// I am not the coordinator, I should send my Gamma matrix out.
shared_ptr <SharedBuffer> buf ( new MemoryBuffer(NULL, sizeof(double) * (d+3) * (d+2) / 2) );
double *Gammabuf = static_cast<double*> (buf->getData());
for(size_t i=0; i<d+2; ++i) {
for(size_t j=0; j<=i; ++j) {
*Gammabuf = Gamma[i][j];
++Gammabuf;
}
}
BufSend(0, buf, query);
return outputArray;
}
else {
// I am the coordinator, I should collect Gamma matrix from workers.
for(InstanceID l = 1; l<query->getInstancesCount(); ++l) {
shared_ptr<SharedBuffer> buf = BufReceive(l, query);
double *Gammabuf = static_cast<double*> (buf->getData());
for(size_t i=0; i<d+2; ++i) {
for(size_t j=0; j<=i; ++j) {
Gamma[i][j] += *Gammabuf;
++Gammabuf;
}
}
}
} // end if getInstanceID() != 0
} //end if InstancesCount() > 1
return writeGamma(d, query);
}
shared_ptr< Array > execute(vector< shared_ptr< Array> >& inputArrays, shared_ptr<Query> query) {
shared_ptr<Array> outputArray(new MemArray(_schema, query));
shared_ptr<Array> inputArray = inputArrays[0];
ArrayDesc inputSchema = inputArray->getArrayDesc();
// Get descriptor of two dimensions d and n.
DimensionDesc dimsN = inputSchema.getDimensions()[0];
DimensionDesc dimsD = inputSchema.getDimensions()[1];
size_t n = dimsN.getCurrLength();
// Note: the input data set should have d+1 dimensions (including Y)
size_t d = dimsD.getCurrLength() - 1;
nlq.N = n;
nlq.d = d;
shared_ptr<ConstArrayIterator> inputArrayIter = inputArray->getConstIterator(0);
Coordinates cellPosition;
size_t i;
double value;
while(! inputArrayIter->end() ) {
shared_ptr<ConstChunkIterator> chunkIter = inputArrayIter->getChunk().getConstIterator();
// For each cell in the current chunk.
// This will skip the empty cells.
while(! chunkIter->end() ) {
cellPosition = chunkIter->getPosition();
value = chunkIter->getItem().getDouble();
nlq.L[ cellPosition[1] ] += value;
nlq.Q[ cellPosition[1] ] += value * value;
++(*chunkIter);
}
++(*inputArrayIter);
}
/**
* The "logical" instance ID of the instance responsible for coordination of query.
* COORDINATOR_INSTANCE if instance execute this query itself.
*/
if(query->getInstancesCount() > 1) {
if(query->getInstanceID() != 0) {
// I am not the coordinator, I should send my Gamma matrix out.
shared_ptr <SharedBuffer> buf ( new MemoryBuffer(NULL, sizeof(double) * (d*2+2) ));
double *Gammabuf = static_cast<double*> (buf->getData());
for(i=1; i<=d+1; ++i) {
*Gammabuf = nlq.L[i];
++Gammabuf;
}
for(i=1; i<=d+1; ++i) {
*Gammabuf = nlq.Q[i];
++Gammabuf;
}
BufSend(0, buf, query);
return outputArray;
}
else {
// I am the coordinator, I should collect Gamma matrix from workers.
for(InstanceID l = 1; l<query->getInstancesCount(); ++l) {
shared_ptr<SharedBuffer> buf = BufReceive(l, query);
double *Gammabuf = static_cast<double*> (buf->getData());
for(i=1; i<=d+1; ++i) {
nlq.L[i] += *Gammabuf;
++Gammabuf;
}
for(i=1; i<=d+1; ++i) {
nlq.Q[i] += *Gammabuf;
++Gammabuf;
}
}
}// end if getInstanceID() != 0
}//end if InstancesCount() > 1
return writeGamma(query);
}
shared_ptr< Array > execute(vector< shared_ptr< Array> >& inputArrays, shared_ptr<Query> query)
{
shared_ptr<Array> outputArray(new MemArray(_schema, query));
shared_ptr<Array> inputArray = inputArrays[0];
ArrayDesc inputSchema = inputArray->getArrayDesc();
// Get descriptor of two dimensions d and n.
DimensionDesc dimsN = inputSchema.getDimensions()[0];
DimensionDesc dimsD = inputSchema.getDimensions()[1];
int64_t n = dimsN.getCurrEnd() - dimsN.getCurrStart() + 1;
// Note: the input data set should have d+1 dimensions (including Y)
d = dimsD.getCurrEnd() - dimsD.getCurrStart();
idY = d+1;
int64_t nStart = dimsN.getCurrStart();
int64_t dStart = dimsD.getCurrStart();
// Get chunk size of n.
int64_t nChunkSize = dimsN.getChunkInterval();
k = ((shared_ptr<OperatorParamPhysicalExpression>&)_parameters[0])->getExpression()->evaluate().getInt64();
if (_parameters.size() == 2) {
idY = ((shared_ptr<OperatorParamPhysicalExpression>&)_parameters[1])->getExpression()->evaluate().getInt64();
}
#ifdef DEBUG
stringstream ss;
ss << getenv("HOME") << "/groupdiagdensegamma-instance-" << query->getInstanceID() << ".log";
log.open(ss.str().c_str(), ios::out);
log << "n = " << n << endl << "d = " << d << endl << "k = " << k << endl;
log << "nStart = " << nStart << endl << "dStart = " << dStart << endl;
log << "nChunkSize = " << nChunkSize << endl;
log << "idY = " << idY << endl;
#endif
shared_ptr<ConstArrayIterator> inputArrayIter = inputArray->getConstIterator(0);
Coordinates chunkPosition;
int64_t i, j, k, m, l;
double value;
NLQ tmp;
map<double, struct NLQ>::iterator it;
while(! inputArrayIter->end() ) {
shared_ptr<ConstChunkIterator> chunkIter = inputArrayIter->getChunk().getConstIterator();
chunkPosition = inputArrayIter->getPosition();
#ifdef DEBUG
log << "Getting into chunk (" << chunkPosition[0] << ", " << chunkPosition[1] << ")." << endl;
#endif
for(i=chunkPosition[0]; i<chunkPosition[0] + nChunkSize; i++) {
if(i == n + nStart) {
#ifdef DEBUG
log << "Reaching row " << i << ", exiting." << endl;
#endif
break;
}
for(j=chunkPosition[1], m=1; j<=chunkPosition[1]+d; j++, m++) {
if(j == d + 1 + dStart) {
#ifdef DEBUG
log << "Reaching column " << j << ", exiting." << endl;
#endif
break;
}
value = chunkIter->getItem().getDouble();
tmp.L[m] = value;
tmp.Q[m] = value * value;
++(*chunkIter);
}
double Y = tmp.L[idY];
it = nlq.find(Y);
if (it == nlq.end()) {
#ifdef DEBUG
log << "Cannot find NLQ entry for class " << Y << ", creating new." << endl;
#endif
nlq[Y].N = 1;
nlq[Y].groupId = Y;
}
else {
nlq[Y].N++;
}
for (k=1, l=1; k<=d+1; k++) {
if (k == idY) {
continue;
}
nlq[Y].L[l] += tmp.L[k];
nlq[Y].Q[l] += tmp.Q[k];
l++;
}
nlq[Y].L[d+1] += tmp.L[idY];
nlq[Y].Q[d+1] += tmp.Q[idY];
}
++(*inputArrayIter);
}
/**
* The "logical" instance ID of the instance responsible for coordination of query.
* COORDINATOR_INSTANCE if instance execute this query itself.
*/
size_t localClassCount = nlq.size();
#ifdef DEBUG
log << "localClassCount = " << localClassCount << endl;
#endif
if(query->getInstancesCount() > 1) {
if(query->getInstanceID() != 0) {
// I am not the coordinator, I should send my NLQ out.
#ifdef DEBUG
log << "I am not the coordinator, I should send my NLQ out." << endl;
#endif
shared_ptr <SharedBuffer> buf ( new MemoryBuffer(NULL, sizeof(struct NLQ) * localClassCount ));
struct NLQ *NLQbuf = static_cast<struct NLQ*> (buf->getData());
for(it = nlq.begin(); it != nlq.end(); it++) {
*NLQbuf = it->second;
++NLQbuf;
}
BufSend(0, buf, query);
#ifdef DEBUG
log << "Exiting." << endl;
#endif
return outputArray;
}
else {
// I am the coordinator, I should collect NLQ from workers.
#ifdef DEBUG
log << "I am the coordinator, I should collect NLQ from workers." << endl;
#endif
for(InstanceID l = 1; l<query->getInstancesCount(); ++l) {
shared_ptr<SharedBuffer> buf = BufReceive(l, query);
if(! buf) {
#ifdef DEBUG
log << "Nothing from instance " << l << ", continue." << endl;
#endif
continue;
}
int64_t remoteClassCount = buf->getSize() / sizeof(struct NLQ);
struct NLQ* NLQbuf = static_cast<struct NLQ*> (buf->getData());
#ifdef DEBUG
log << "Received " << remoteClassCount << " entries from instance " << l << endl;
#endif
for(i=0; i<remoteClassCount; ++i) {
it = nlq.find(NLQbuf->groupId);
if( it == nlq.end() ) {
#ifdef DEBUG
log << "Cannot find NLQ entry for class " << NLQbuf->groupId << ", creating new." << endl;
#endif
nlq[NLQbuf->groupId] = *NLQbuf;
}
else {
it->second.N += NLQbuf->N;
for(j=1; j<=d+1; ++j) {
it->second.L[j] += NLQbuf->L[j];
it->second.Q[j] += NLQbuf->Q[j];
}
}
++NLQbuf;
}
#ifdef DEBUG
log << "Merge complete." << endl;
#endif
}
}// end if getInstanceID() != 0
}//end if InstancesCount() > 1
return writeGamma(query);
}
/**
* Perform operator-specific checks of input and return the shape of the output. Currently,
* the output array must exist.
* @param schemas the shapes of the input arrays
* @param query the query context
*/
ArrayDesc inferSchema(std::vector< ArrayDesc> schemas, shared_ptr< Query> query)
{
assert(schemas.size() == 1);
assert(_parameters.size() == 1);
string arrayName = ((shared_ptr<OperatorParamReference>&)_parameters[0])->getObjectName();
ArrayDesc const& srcDesc = schemas[0];
//Ensure attributes names uniqueness.
ArrayDesc dstDesc;
if (!SystemCatalog::getInstance()->getArrayDesc(arrayName, dstDesc, false))
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_ARRAY_DOESNT_EXIST) << arrayName;
}
if(dstDesc.isImmutable())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_ILLEGAL_OPERATION)
<< "Target of INSERT must be a mutable array";
}
Dimensions const& srcDims = srcDesc.getDimensions();
Dimensions const& dstDims = dstDesc.getDimensions();
if (srcDims.size() != dstDims.size())
{
//TODO: this will get lifted when we allow redimension+insert in the same op
//and when we DO implement redimension+insert - we will need to match attributes/dimensions by name, not position.
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_ILLEGAL_OPERATION)
<< "Temporary restriction: target of INSERT must have same dimensions as the source";
}
for (size_t i = 0, n = srcDims.size(); i < n; i++)
{
if( srcDims[i].getType() != TID_INT64 || dstDims[i].getType() != TID_INT64)
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_ILLEGAL_OPERATION)
<< "Temporary restriction: INSERT only supports integer dimensions";
}
//TODO: we can also allow arrays that are smaller whose length is not evenly divided by chunk interval
//but then we have to detect "edge chunks" and rewrite them cleverly
if( srcDims[i].getStartMin() != dstDims[i].getStartMin() ||
srcDims[i].getChunkInterval() != dstDims[i].getChunkInterval() ||
srcDims[i].getChunkOverlap() != dstDims[i].getChunkOverlap() ||
srcDims[i].getEndMax() > dstDims[i].getEndMax() ||
( srcDims[i].getEndMax() < dstDims[i].getEndMax() &&
srcDims[i].getLength() % srcDims[i].getChunkInterval() != 0))
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_DIMENSIONS_DONT_MATCH)
<< srcDims[i].getBaseName() << dstDims[i].getBaseName();
}
}
Attributes const& srcAttrs = srcDesc.getAttributes(true);
Attributes const& dstAttrs = dstDesc.getAttributes(true);
if (srcAttrs.size() != dstAttrs.size())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_ILLEGAL_OPERATION)
<< "Temporary restriction: target of INSERT must have same attributes as the source";
}
for (size_t i = 0, n = srcAttrs.size(); i < n; i++)
{
if(srcAttrs[i].getType() != dstAttrs[i].getType())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_ATTRIBUTE_TYPE)
<< srcAttrs[i].getName() << srcAttrs[i].getType() << dstAttrs[i].getType();
}
//can't store nulls into a non-nullable attribute
if(!dstAttrs[i].isNullable() && srcAttrs[i].isNullable())
{
throw USER_EXCEPTION(SCIDB_SE_INFER_SCHEMA, SCIDB_LE_WRONG_ATTRIBUTE_FLAGS)
<< srcAttrs[i].getName();
}
}
//Note: let us NOT add arrayID numbers to the schema - because we do not have our ArrayID yet.
//We will get our ArrayID when we execute and create the array. Until then - don't bother.
//Old store code adds the arrayID to the schema - but that's the arrayID of the previous version,
//not the new version created by the op. A dangerous fallacy - stupid and unnecessary.
return ArrayDesc(arrayName, dstDesc.getAttributes(), dstDesc.getDimensions(), dstDesc.getFlags());
}
