IMP Manual
for IMP version 2.15.0
|
The IMP code is automatically tested for consistency and stability in a number of ways.
IMP is built nightly from source code on a variety of different machines in the Sali lab, and in several different configurations, and all the test cases, examples, and benchmarks run. This aims to detect problems with new code relatively quickly. See the nightly build status page for latest results. This is done for the head of the develop
branch of the code in the GitHub repository. When a new stable release is in preparation, this is also done for the corresponding release/x.y.z
branch and also for the main
branch once the new release is merged there. Results for these branches can also be viewed at the nightly build state page by selecting the branch from the dropdown at the top left of the page.
A subset of the IMP code is built using Travis CI every time code is pushed to the develop
branch on GitHub. (Only a subset is built because building all of IMP would exceed Travis CI's time limits.) This is helpful to quickly detect compilation errors.
The entire IMP code is periodically subjected to static code analysis using Coverity. This is useful to detect some types of coding errors, such as buffer overflows.
The Coverity scan is not currently completely automated, since the compilation is quite expensive and Coverity imposes strict limits on the frequency of builds. It is currently run manually on an older Mac running OS X 10.6 (the newer Macs in the Sali lab use the clang
compiler, which Coverity doesn't support, and the build takes a really long time on a Linux box).
Note that Coverity currently reports a rather high defect density for IMP. Many of these are false positives - for example, Coverity complains about all of the geometric primitives not being initialized, but this is intentional.