xlapack.solveCholesky test precision too high for 32 bit machines about xtensor-blas HOT 7 OPEN

Thanks for the report.

We would be interested in a fix, but presumably it would be interesting to include an i386 test in the CI if we want to catch these.

from xtensor-blas.

A pragmatic workaround could be to increase the test tolerance e.g. from 1e-18 to 1e-16 (as an example. I'm not certain what the effective tolerance here is). That could be done here using EXPECT_NEAR in place of EXPECT_DOUBLE_EQ.

from xtensor-blas.

Thanks!
Would you like to open a pull request?

from xtensor-blas.

I can do it. I'll need to investigate what tolerance i386 needs.
Well, I guess it's just 0.13757507429403265 - 0.13757507429403248 = 2e-16. (the effective tolerance of EXPECT_DOUBLE_EQ must be something like 1e-16)

from xtensor-blas.

Also affects armhf, confirming it's a 32-bit issue.
https://ci.debian.net/data/autopkgtest/testing/armhf/x/xtensor-blas/15977642/log.gz

from xtensor-blas.

armhf is a little coarser,
0.13757507429403265 - 0.13757507429403243 = 2.220446049250313e-16

from xtensor-blas.

Looks like 5e-16 is enough. i386 and armhf are both passing now with PR #212,
https://ci.debian.net/data/autopkgtest/testing/i386/x/xtensor-blas/15983027/log.gz
https://ci.debian.net/data/autopkgtest/testing/armhf/x/xtensor-blas/15982850/log.gz

Eventually test logs will appear at https://ci.debian.net/packages/x/xtensor-blas/

from xtensor-blas.