I built atlas-lapack from source to maximize my hardware capabilities. This might also work for regular ATLAS. I have not tested with it. I did not use OpenBLAS at all.

OpenCVFindLAPACK.cmake needs to be edited because of C++ name mangling which prevents the linker from finding the appropriate ATLAS symbols.

Edit OpenCVFindLAPACK.cmake

Change:

if(NOT " ${__content_str}" STREQUAL " ${_lapack_include_str}")
file(WRITE "${CBLAS_H_PROXY_PATH}" "${_lapack_include_str}")
endif()

To:

if(NOT " ${__content_str}" STREQUAL " ${_lapack_include_str}")
file(WRITE "${CBLAS_H_PROXY_PATH}" "extern 'C' { ${_lapack_include_str} }")
endif()

#############################################################################

Here is a known good build configuration which worked for me.

OPENCV_VERSION="3.3.1"
PYTHON_VERSION="3.6.3"

CFLAGS="-march=native -O2 -pipe" \
CXXFLAGS="-march=native -O2 -pipe" \
cmake \
-D BUILD_TESTS=OFF \
-D ENABLE_PRECOMPILED_HEADERS=OFF \
-D CMAKE_INSTALL_PREFIX="/opt/opencv_${OPENCV_VERSION}_python_${PYTHON_VERSION}" \
-D OPENCV_EXTRA_MODULES_PATH="../../opencv_contrib-3.3.1/modules" \
-D BUILD_OPENCV_PYTHON2=OFF \
-D BUILD_OPENCV_PYTHON3=ON \
-D CUDA_NVCC_FLAGS="--expt-relaxed-constexpr -ccbin gcc-6" \
-D CUDA_GENERATION=Pascal \
-D WITH_CUDA=ON \
-D ENABLE_FAST_MATH=1 \
-D CUDA_FAST_MATH=1 \
-D WITH_CUBLAS=1 \
-D WITH_EIGEN=ON \
-D BUILD_TIFF=ON \
-D BUILD_OPENCV_JAVA=OFF \
-D BUILD_PERF_TESTS=OFF \
-D CMAKE_BUILD_TYPE=RELEASE \
-D PYTHON3_LIBRARY=$(python -c "import os.path; from distutils.sysconfig import get_config_var; print(os.path.join(get_config_var('LIBDIR'), get_config_var('LDLIBRARY')))") \
-D PYTHON3_EXECUTABLE=$(which python) \
-D PYTHON3_INCLUDE_DIR=$(python -c "from distutils.sysconfig import get_python_inc; print(get_python_inc())") \
-D PYTHON3_PACKAGES_PATH=$(python -c "from distutils.sysconfig import get_python_lib; print(get_python_lib())") ..