search

coneoproject / COFFEE

COFFEE - A COmpiler For Fast Expression Evaluation
Other
9 stars 6 forks source link

potential bug with compiler setting #120

Open tj-sun opened 7 years ago

tj-sun commented 7 years ago

It appears the following test passes with gcc but fails with icc regression/test_helmholtz_sphere.py::test_helmholtz_sphere_lowestorder[MeshClass1]

wence- commented 7 years ago

fails in what way?

tj-sun commented 7 years ago

convergence rate

assert (l2conv > 1.7).all() E assert <built-in method all of numpy.ndarray object at 0x7f88b791cd00>() E + where <built-in method all of numpy.ndarray object at 0x7f88b791cd00> = array([ 0.04270851, 0.01346899]) > 1.7.all

wence- commented 7 years ago

OK, yes, that's bad.

miklos1 commented 7 years ago

Is this a bug in the Intel compiler, TSFC, or COFFEE?

tj-sun commented 7 years ago

On first look it appears that 1) the c files are identical 2) the compilation commands are identical So I guess it's probably problem with either the compiler (unlikely?) or that we ought to use different flags for different compilers somehow.

Will have another look later.

miklos1 commented 7 years ago

Before saying that it's a C compiler bug, you must judge that the generated code is standard conforming and does not contain undefined behaviour.

Another thing to try with potential C compiler bugs is the change/lower the optimisation level.