Travis CI rebased

[jeffdonahue]

Hi, I've taken PR #667 (thanks @huyng!) and rebased it onto some commits cutting down on the duplication between CPU and GPU tests to make it easier and more reliable to run only the CPU tests, and cut down on duplicated code throughout the test suite. I added a make runtestnogpu target to run only the tests that do not use the GPU (had to append CPUGPU to some tests that use both CPU and GPU and can't be separated).

I made a few changes to the .travis.yml in d95aea8 for CPU-only testing and notification by email only going to the committer. Others can see build status on the PR or travis feed.

[jeffdonahue]

whoohoo, PR integration working (and build passed!)

[shelhamer]

I'm so excited for this! Please squash before merge.

[shelhamer]

Why HDR? Everywhere else is HXX.

[shelhamer]

Nice extension of the testing types and arranging into include/caffe.

[jeffdonahue]

Thanks for taking a look @shelhamer! Squashed history, changed TEST_HDRS to TEST_HXX_SRCS in Makefile, and removed CUSTOM_CXX.

[shelhamer]

Sidenote: much of the test time is downloading and installing the CUDA requirements plus doing the nvcc compilation, so this should act more lively once the device abstraction and CPU-only build are done.

[shelhamer]

[jeffdonahue]

sweet, thanks for the quick merge

[shelhamer]

Wait, I remember why I did CUSTOM_CXX now. At least on OS X the CXX ?= method breaks because it picks up the value c++ from somewhere even though it's not in my environment as reported by env.

@jeffdonahue can you revert the change to the Makefile to bring back the CUSTOM_CXX logic and make Travis set the set CUSTOM_CXX (e.g. sed the Makefile.config or call CUSTOM_CXX=g++ make instead)?

[longjon]

@shelhamer, right, I believe CXX and a few others are set to defaults by
make itself for use in implicit rules, see
https://www.gnu.org/software/make/manual/html_node/Implicit-Variables.html.
On Jul 14, 2014 6:58 AM, "Evan Shelhamer" notifications@github.com wrote:

Wait, I remember why I did CUSTOM_CXX now. At least on OS X the CXX ?=
method breaks because it picks up the value c++ from somewhere even
though it's not in my environment as reported by env.

Can you revert the change to CXX and make Travis set the Makefile.config
CUSTOM_CXX field instead (e.g. run a sed command or whatever)?

—
Reply to this email directly or view it on GitHub
#681 (comment).

[jeffdonahue]

Re-added CUSTOM_CXX stuff to Makefile and Makefile.config.example in #694. I didn't change the travis build logic yet to use it yet because I'm not sure how to get the separate, per-compiler, builds behavior working using the explicit env variable export or Makefile.config change, rather than the Travis compiler field (feel free to take this on if you'd like). I'm not sure about make itself setting CXX and other variables; when I had the clang build enabled (can try yourself by uncommenting the - clang line in .travis.yml) it was failing so clearly the CXX setting was having some effect.

[longjon]

@jeffdonahue travis sets the CXX environment variable, which make picks up to set its own CXX.

Witness:

$ echo $CXX

$ cat > Makefile
all:  
    echo $(CXX)

$ make
echo g++
g++

$ CXX=clang++ make
echo clang++
clang++

$ cat > Makefile
CXX ?= clang++
all:
    echo $(CXX)

$ make
echo g++
g++

See https://www.gnu.org/software/make/manual/html_node/Environment.html#Environment.

[jeffdonahue]

I see -- so make lets the CXX env var setting override its default, but specifying CXX as a Makefile variable doesn't affect anything because it just keeps the default setting when it runs into a line that tries to set it to something else? Interesting...

[longjon]

You can override the value with a :=, but if you try to specify it with a ?= line, yes. In effect make imports your environment and then has its own ?= lines for a bunch of default variables.

[jeffdonahue]

Oh okay, that makes sense -- thanks for the explanation.