This change removes the cudaDevAttrIntegrated check and only requires
pageable memory access. This change allows to run GPU accelerated
channel convolution on GH machines.