[SYSTEMDS-1817] Full standard deviation fails due to wrong data type

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: SystemML 0.15
• Assignee: Matthias Boehm
• Reporter: Matthias Boehm
• Created: 29/Jul/17 06:13
• Updated: 09/Sep/17 03:32
• Resolved: 29/Jul/17 22:51

Description

The following script currently fails due to incorrect hop construction of sd(Xi) at language level:

X = rand(rows=1000, cols=100);
R = matrix(0, ncol(X), 4)
parfor(i in 1:ncol(X), log=DEBUG) {
   Xi = X[,i];
   R[i,1] = mean(Xi);                  # mean
   R[i,2] = sd(Xi);                    # standard deviation
   R[i,3] = moment(Xi,3)/(sd(Xi)^3);   # skewness
   R[i,4] = moment(Xi,4)/(sd(Xi)^4)-3; # kurtosis
}

The stacktrace is as follows

Caused by: org.apache.sysml.runtime.DMLRuntimeException: Variable '_Var1703' does not exist in the symbol table.
	at org.apache.sysml.runtime.controlprogram.context.ExecutionContext.getMatrixObject(ExecutionContext.java:186)
	at org.apache.sysml.runtime.controlprogram.context.ExecutionContext.setMatrixOutput(ExecutionContext.java:451)
	at org.apache.sysml.runtime.instructions.cp.MatrixBuiltinCPInstruction.processInstruction(MatrixBuiltinCPInstruction.java:59)
	at org.apache.sysml.runtime.controlprogram.ProgramBlock.executeSingleInstruction(ProgramBlock.java:286)
	... 6 more

which is cause by the following instruction which has an output type MATRIX although it should be SCALAR.

--------CP uavar _mVar3.MATRIX.DOUBLE _mVar5.MATRIX.DOUBLE 24

Interestingly, this only shows up when there are multiple consumers with matrix and scalar data types.