relax tolerance in the xlapack.solveCholesky test #212
Conversation
The xlapack.solveCholesky test in test_lapack.cpp fails on 32-bit architectures (i386/i686):
[ RUN ] xlapack.solveCholesky
/tmp/autopkgtest-lxc.gg3nslld/downtmp/autopkgtest_tmp/test_lapack.cpp:166: Failure
Expected equality of these values:
x_expected[i]
Which is: 0.13757507429403265
x[i]
Which is: 0.13757507429403248
[ FAILED ] xlapack.solveCholesky (0 ms)
This patch relaxes test tolerance by using EXPECT_NEAR with abstol=2e-16 instead of EXPECT_DOUBLE_EQ.
Fixes Issue xtensor-stack#211.
|
Don't merge this PR until we confirm it fixes the tests on 32-bit arches. |
The armhf discrepancy is coarser, 0.13757507429403265 - 0.13757507429403243 = 2.220446049250313e-16 So relax test tolerance to 5e-16.
|
Curiously, applying tolerance 2e-16 gives this warning According to that, the i386 test should have been within tolerance in the first place (0.13757507429403265 - 0.13757507429403248 = 1.6653e-16). This patch is now set to 5e-16, we'll know soon if that's relaxed enough. |
Let me know if that's fine with the 32 bits testing you got in Debian! |
|
Looks like 5e-16 is enough. i386 and armhf are both passing now, Eventually test logs will appear at https://ci.debian.net/packages/x/xtensor-blas/ Thumbs up from me for the PR now. |
The xlapack.solveCholesky test in test_lapack.cpp fails on 32-bit architectures (i386/i686):
[ RUN ] xlapack.solveCholesky
/tmp/autopkgtest-lxc.gg3nslld/downtmp/autopkgtest_tmp/test_lapack.cpp:166: Failure
Expected equality of these values:
x_expected[i]
Which is: 0.13757507429403265
x[i]
Which is: 0.13757507429403248
[ FAILED ] xlapack.solveCholesky (0 ms)
This patch relaxes test tolerance by using EXPECT_NEAR with abstol=2e-16 instead of EXPECT_DOUBLE_EQ.
Fixes Issue #211.