<div dir="ltr"><div>So far I could not reproduce the behaviour with any other system than the one provided by Andres. I tried <br></div><div> - QS/regtest-admm-1/CH3-BP-NONE.inp</div><div> - QS/regtest-hybrid-4/CH4-PBE0_TC_LRC.inp</div><div> - the Si-8 example from <a href="https://www.cp2k.org/howto:static_calculation">https://www.cp2k.org/howto:static_calculation</a></div><div> - the Si-8 example with 2 Si replaced by Al</div><div>I took the subsys from these 4 examples and replaced the one in the input from Andres. All of them behave as they should with CUDA+OMP.</div><div><br></div><div>So it seems to be difficult to catch and maybe does not trigger very often?</div><div>That is probably why it does not show up in the regtests.<br></div><div><br></div><div>Best,</div><div>Fabian<br></div><br>On Monday, 4 May 2020 18:26:40 UTC+2, Ole Schütt wrote:<blockquote class="gmail_quote" style="margin: 0;margin-left: 0.8ex;border-left: 1px #ccc solid;padding-left: 1ex;">Hi Leopold,
<br>
<br>I agree that this is a serious issue that should have been caught by our
<br>testing.
<br>AFAIK, the performance tests on the dashboard do not check the computed
<br>results.
<br>
<br>However, we do have a daily CUDA regtest:
<br><a href="https://dashboard.cp2k.org/archive/cuda-pascal/index.html" target="_blank" rel="nofollow" onmousedown="this.href='https://www.google.com/url?q\x3dhttps%3A%2F%2Fdashboard.cp2k.org%2Farchive%2Fcuda-pascal%2Findex.html\x26sa\x3dD\x26sntz\x3d1\x26usg\x3dAFQjCNGbB_6LENymlxdMl98matMHdziP3A';return true;" onclick="this.href='https://www.google.com/url?q\x3dhttps%3A%2F%2Fdashboard.cp2k.org%2Farchive%2Fcuda-pascal%2Findex.html\x26sa\x3dD\x26sntz\x3d1\x26usg\x3dAFQjCNGbB_6LENymlxdMl98matMHdziP3A';return true;">https://dashboard.cp2k.org/<wbr>archive/cuda-pascal/index.html</a>
<br>It uses two OpenMP threads, which might not be enough to exceed the
<br>tolerance thresholds?
<br>
<br>-Ole
<br></blockquote></div>