Dear Robert,
try setting these MKL environment variables
export MKL_INTERFACE_LAYER=GNU,LP64
export MKL_THREADING_LAYER=GNU
the example should then report correct results (and regardless of the
version of R). There is no need to rebuild R from source, but when you
are building from source, this problem would be detected by checks in R
configure scripts. I can reproduce the problem on Ubuntu 19.04 and MKL
installed from binaries as described in Dirk's blog, both with R 3.5.2
and R-devel, and the problem goes away on both when the variables are set.
Best
Tomas
On 7/17/19 3:18 PM, Robert B. Gramacy wrote:
Dear R-devel team,
I've encountered problems with recent Rs (>= 3.5.3) and MKL. I've
instructions and many versions of MKL. All works fine in my Ubuntu 18
setup for R 3.5.2 and older. Carrying out the install and linking
instructions works fine for later Rs, but calculations involving linear
algebra become corrupted. E.g., following Dirk's very simple svd example
from that link above ...
n <- 1e3
X <- matrix(rnorm(n*n),n,n)
Xsvd <- svd(X)
Xsvd$d[1:5]
[1] 63.09366 62.47868 62.26147 61.94530 61.84727
... you get Eigenvalues that are bigger than 1e26 in R-devel, say, but
(correct) values near 60 with R 3.5.2.
I just wanted to throw that out there in case anyone else had come across
something similar. I don't mind using a slightly older R/MKL for my
stuff. I might speculate that the Microsoft R Open folks are having
similar issues because they're still using older Rs.
I haven't tried compiling and linking from source (
https://cran.r-project.org/doc/manuals/R-admin.html#MKL) because I
had success with that in some time.
-B
[[alternative HTML version deleted]]