Compile errors in boost-eigen branch

[alito]

The boost-eigen branch does not compile for me. This is after the merge from kloudkl (8d894f0) (I never tested it before that).

boost::math::nextafter should only be parametrised on one type (or so says gcc here, and what I read on the boost docs).

Also, the tests for test_random_number_generator reference mean_bound and sample_mean but they need to be referenced as this->mean_bound (again, that's just what my gcc thinks).

This branch contains those two fixes. After this, everything compiles and the tests run fine (except for some GPU tests)

[dnouri]

Sorry, it must have been that I tested an older version of kloudkl's branch when I confirmed it was working. :-(

I can anyway now confirm that merging alito's branch into boost-eigen allows me to compile and run caffe again. I've run all tests and two of them are failing:

• "PoolingLayerTest/1.TestGPUGradientAve" errors with "pooling_layer.cu:309] Cuda kernel failed. Error: too many resources requested for launch"
• "LRNLayerTest/0.TestGPUGradient" errors with "lrn_layer.cu:80] Cuda kernel failed. Error: too many resources requested for launch"

I haven't tested against the main branch, so can't say if these occur there as well.

[jeffdonahue]

Thanks @alito! We're confirming this builds and passes with both Boost 1.46 and Boost 1.54. (kloudkl's version compiles with Boost 1.54, but not 1.46.)

[kloudkl]

All the tests passed under the environment as follows.
gcc (Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3
Boost 1.54/1.55 installed from source
NVIDIA GTX 560 Ti with 1GB memory)
CUDA 5.5

Is it necessary to setup a continuous integration infrastructure like the Chromium build infrastructure to test various possible combinations?