This give ~4.5% speed up for unrolled version. Less for other version.

in ConvOp.__init__, if the unroll value are not good, we try to find a good one. Otherwise we set 1 and print an OPTIMISATION WARNING.

We do this as it is probably safer to use default value for the optimisation parameter if they are bad and execute then to fail.

Print the time in second and the cumulative time too.

small bugfix. The ConvOp was not unrolled when only the unroll_batch or the unroll_kern option was specified.

check that when we unroll, we unroll with a minimim value of 1. removed the two useless generation fct that generate the unroll only by batch or by kern. We have a fct that unroll by both of them. Set the unroll value to one to have the equivalent fct.

added a new unrolled verion of ConvOp(not used by default). It unroll the batch and the kernel at the same time. This give the biggest speed up in my test. Also modified the test_multilayer_conf fct to allow testing the different parameter for the unroll.

added the unroll version by kern. If you put the ConvOp option unroll_batch or unroll_kern>0, we will those version with the provided unrool size.

added option to ConvOp that allow using the unroll version of the code. The default is to don't use this version of the code.

added a function to generate an unrolled version of the c code for the ConvOp. The code is not used.

scipy.signal.convolve2d causes errors on atexit. i moved imports to make it possible to use our convolve op without triggering this problem

self). Also did some housecleaning