tskisner / hpcports Goto Github PK

On OS X Lion:

gcc -O2 -g -m64 -fPIC -I. -o imsetwcs.o -c imsetwcs.c
imsetwcs.c:1225: internal compiler error: Segmentation fault: 11
Please submit a full bug report,
with preprocessed source if appropriate.
See URL:http://developer.apple.com/bugreporter for instructions.

The sqlite3 search paths in the python setup.py file sometimes pick up an incompatible sqlite installation. This setup.py should be patched to use only the hpcports sqlite3 path (built-in or overridden).

Currently hpcports uses an unreleased (and ancient) version of healpix that was autotoolized by me for easy installation. In order to update to the latest upstream sources, we should split this into C, Fortran, and C++ packages.

When building with Intel compiler and static MKL, there are circular dependencies. This requires --preserve-dup-deps option to libtool. Some packages don't seem to pass this option on, even when it is specified in LDFLAGS. This is the current holdup for Intel compiled hpcports on edison.nersc.gov.

I'm working on cori.

The current galsim is 1.3. I would like to go up to the latest pre-release tag. The tag is v1.5.0-alpha

I've been using this code for quite some time and it is actually quite stable. And I need the functionality for the work I'm planning to do.

Is it possible to make this pre-release available, but not as default? And then when the full 1.5 release appears move over to that as the default?

If the variable _PREFIX has been overridden, do not clobber this when generating the module file stub for the overridden package.

These are now distributed separately from the C library

We should add TARGET / HOST variables to the system config files, and (if not empty) pass these to autoconf-based packages to disable execution of configure tests. CMAKE packages need more investigation. Cross compile is needed for several machines (BGQ, XK7, MIC systems, etc).

A strange difference between Edison and Cori: on Cori, matplotilb doesn't support jpeg-format output, but on Edison it does.

Overridden package variables are specified in the per-system make file snippets. These are used to build module files for the overridden packages which prepend to CPATH, LIBRARY_PATH, etc and propagate any variables in the system make file. The internal hpcports build system does not use these modules, instead relying on the per-package shell snippets which are sourced. No such shell snippets exist yet for overridden packages.

When working on standard linux systems and overriding many packages, it is tedious to specify all the variables. Should we switch to explicitly enabling packages? Should we have standard override "profiles" for Ubuntu, RedHat, OS X with Homebrew, etc?

On some systems (Cray), the build system is slow. Not sure if this is due to the large number of rules that control package dependencies, or a large environment size when spawning a new shell. Fixing this may require switching to a standalone package tool (like apt) rather than using Make.

The pull request at astropy/astropy#2816 allows one to correctly query the SFD98 map files with coordinate transformation routines in astropy.

This PR wasn't included in the current hpcport version of astropy.

It would be nice to have a mechanism that allows multiple packages to "provide" a virtual package for purposes of meeting dependencies. This feature exists in other package managers (apt, macports, etc).

More documentation about all options available in package file creation.

For all compilers we care about (GCC, Intel, IBM), we can use C_INCLUDE_PATH and CPLUS_INCLUDE_PATH to specify directories to search for headers. Add this to modulefiles and package shell snippets. Need to find equivalent variables for fortran module search.