Imran Younus created SYSTEMML-1167:
--------------------------------------
Summary: problem with matrix indexing
Key: SYSTEMML-1167
URL: https://issues.apache.org/jira/browse/SYSTEMML-1167
Project: SystemML
Issue Type: Bug
Components: Compiler, Runtime
Reporter: Imran Younus
I've created a recursive QR decomposition algorithm (see
https://github.com/apache/incubator-systemml/pull/322). As part of this
algorithm I've implemented a function that converts householder matrix to Q:
{code}
QfromH = function(Matrix[double] H)
return(Matrix[double] Q) {
m = nrow(H);
n = ncol(H);
ones = matrix(1, m, 1);
eye = diag(ones);
Q = eye[,1:n];
for (j in n:1) {
v = H[j:m,j]
b = as.scalar(2/(t(v) %*% v))
Q[j:m, j:n] = Q[j:m, j:n] - (b * v) %*% (t(v) %*% Q[j:m, j:n])
}
}
{code}
In this code, I'm initializing {{Q}} from {{eye\[,1:n\]}}. So {{Q}} should be a
{{m x n}} matrix with {{m > n}}. This function is called in the base case for
the recursive QR algorithm.
At the first call, the matrix {{Q}} has correct shape. But, in every successive
call it become a {{m x m}} matrix instead of {{m x n}} matrix and the algorithm
breaks down.
It turns out that if the compiler optimization level is set to 0, then this
problem goes away. Please see the discussion at
https://github.com/apache/incubator-systemml/pull/322
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
