From c8a062b6d439ab51a216e2522f4651a802343d90 Mon Sep 17 00:00:00 2001
From: Berk Hess <hess@kth.se>
Date: Mon, 7 Oct 2019 11:59:56 +0200
Subject: [PATCH] Make nbnxm a proper module

Added definition of module_nbnxm to doxygen documentation.
Required removing the path from many #includes inside
the module.

Change-Id: I1d46aefaa83b0be674c14a4ef5ee16573eee0e1a
---
 docs/doxygen/lib/nbnxm.md                     |  30 +++++
 docs/doxygen/user/mainpage.md                 |   2 +
 src/gromacs/mdlib/sim_util.cpp                |   1 -
 src/gromacs/nbnxm/atomdata.cpp                |   4 +-
 src/gromacs/nbnxm/clusterdistancekerneltype.h |   2 +-
 src/gromacs/nbnxm/cuda/nbnxm_cuda.cu          |  22 ++--
 src/gromacs/nbnxm/cuda/nbnxm_cuda.h           |   2 +
 .../nbnxm/cuda/nbnxm_cuda_kernel_F_noprune.cu |   6 +-
 .../nbnxm/cuda/nbnxm_cuda_kernel_F_prune.cu   |   6 +-
 .../cuda/nbnxm_cuda_kernel_VF_noprune.cu      |   6 +-
 .../nbnxm/cuda/nbnxm_cuda_kernel_VF_prune.cu  |   6 +-
 .../cuda/nbnxm_cuda_kernel_pruneonly.cuh      |   2 +-
 src/gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh |  84 ++++++------
 src/gromacs/nbnxm/gpu_jit_support.h           |   2 +
 src/gromacs/nbnxm/gpu_types_common.h          |   2 +-
 src/gromacs/nbnxm/grid.cpp                    |   4 +-
 src/gromacs/nbnxm/gridset.cpp                 |   3 +-
 src/gromacs/nbnxm/kernel_common.cpp           |   1 +
 src/gromacs/nbnxm/kernel_common.h             |   6 +-
 .../kernel_simd_2xmm_kernel.cpp.pre           |   4 +-
 .../kernel_simd_4xm_kernel.cpp.pre            |   4 +-
 .../make_verlet_simd_kernel_files.py          |   2 +-
 src/gromacs/nbnxm/kerneldispatch.cpp          |  10 +-
 .../nbnxm/kernels_reference/kernel_ref.cpp    |  30 ++---
 .../kernels_reference/kernel_ref_includes.h   |   6 +-
 .../kernels_reference/kernel_ref_outer.h      |  12 +-
 .../kernels_reference/kernel_ref_prune.h      |   1 +
 .../kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp  |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp |   6 +-
 ...rnel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp    |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp   |   6 +-
 ...kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp |   6 +-
 ...kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp |   6 +-
 ...ernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp |   6 +-
 ...el_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJFSw_F.cpp       |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp      |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp   |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJPSw_F.cpp       |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp      |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp   |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJ_F.cpp          |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJ_VF.cpp         |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp      |   6 +-
 .../kernel_ElecEw_VdwLJCombGeom_F.cpp         |   6 +-
 .../kernel_ElecEw_VdwLJCombGeom_VF.cpp        |   6 +-
 .../kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp     |   6 +-
 .../kernel_ElecEw_VdwLJCombLB_F.cpp           |   6 +-
 .../kernel_ElecEw_VdwLJCombLB_VF.cpp          |   6 +-
 .../kernel_ElecEw_VdwLJCombLB_VgrpF.cpp       |   6 +-
 .../kernel_ElecEw_VdwLJEwCombGeom_F.cpp       |   6 +-
 .../kernel_ElecEw_VdwLJEwCombGeom_VF.cpp      |   6 +-
 .../kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp   |   6 +-
 .../kernel_ElecEw_VdwLJFSw_F.cpp              |   6 +-
 .../kernel_ElecEw_VdwLJFSw_VF.cpp             |   6 +-
 .../kernel_ElecEw_VdwLJFSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecEw_VdwLJPSw_F.cpp              |   6 +-
 .../kernel_ElecEw_VdwLJPSw_VF.cpp             |   6 +-
 .../kernel_ElecEw_VdwLJPSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecEw_VdwLJ_F.cpp                 |   6 +-
 .../kernel_ElecEw_VdwLJ_VF.cpp                |   6 +-
 .../kernel_ElecEw_VdwLJ_VgrpF.cpp             |   6 +-
 ...ernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp |   6 +-
 ...rnel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp |   6 +-
 ...l_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp |   6 +-
 ...kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp |   6 +-
 ...nel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp |   6 +-
 ...nel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp |   6 +-
 ...el_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp |   6 +-
 ...ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp    |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp   |   6 +-
 ...kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp    |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp   |   6 +-
 ...kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJ_F.cpp       |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp      |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp   |   6 +-
 .../kernel_ElecQSTab_VdwLJCombGeom_F.cpp      |   6 +-
 .../kernel_ElecQSTab_VdwLJCombGeom_VF.cpp     |   6 +-
 .../kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp  |   6 +-
 .../kernel_ElecQSTab_VdwLJCombLB_F.cpp        |   6 +-
 .../kernel_ElecQSTab_VdwLJCombLB_VF.cpp       |   6 +-
 .../kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp    |   6 +-
 .../kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp    |   6 +-
 .../kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp   |   6 +-
 ...kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTab_VdwLJFSw_F.cpp           |   6 +-
 .../kernel_ElecQSTab_VdwLJFSw_VF.cpp          |   6 +-
 .../kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp       |   6 +-
 .../kernel_ElecQSTab_VdwLJPSw_F.cpp           |   6 +-
 .../kernel_ElecQSTab_VdwLJPSw_VF.cpp          |   6 +-
 .../kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp       |   6 +-
 .../kernel_ElecQSTab_VdwLJ_F.cpp              |   6 +-
 .../kernel_ElecQSTab_VdwLJ_VF.cpp             |   6 +-
 .../kernel_ElecQSTab_VdwLJ_VgrpF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJCombGeom_F.cpp         |   6 +-
 .../kernel_ElecRF_VdwLJCombGeom_VF.cpp        |   6 +-
 .../kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp     |   6 +-
 .../kernel_ElecRF_VdwLJCombLB_F.cpp           |   6 +-
 .../kernel_ElecRF_VdwLJCombLB_VF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJCombLB_VgrpF.cpp       |   6 +-
 .../kernel_ElecRF_VdwLJEwCombGeom_F.cpp       |   6 +-
 .../kernel_ElecRF_VdwLJEwCombGeom_VF.cpp      |   6 +-
 .../kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp   |   6 +-
 .../kernel_ElecRF_VdwLJFSw_F.cpp              |   6 +-
 .../kernel_ElecRF_VdwLJFSw_VF.cpp             |   6 +-
 .../kernel_ElecRF_VdwLJFSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJPSw_F.cpp              |   6 +-
 .../kernel_ElecRF_VdwLJPSw_VF.cpp             |   6 +-
 .../kernel_ElecRF_VdwLJPSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJ_F.cpp                 |   6 +-
 .../kernel_ElecRF_VdwLJ_VF.cpp                |   6 +-
 .../kernel_ElecRF_VdwLJ_VgrpF.cpp             |   6 +-
 .../nbnxm/kernels_simd_2xmm/kernel_outer.h    |  12 +-
 .../nbnxm/kernels_simd_2xmm/kernel_prune.cpp  |   2 +-
 .../nbnxm/kernels_simd_2xmm/kernel_prune.h    |   1 +
 src/gromacs/nbnxm/kernels_simd_2xmm/kernels.h | 120 +++++++++---------
 .../kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp  |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp |   6 +-
 ...rnel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp    |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp   |   6 +-
 ...kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp |   6 +-
 ...kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp |   6 +-
 ...ernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp |   6 +-
 ...el_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJFSw_F.cpp       |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp      |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp   |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJPSw_F.cpp       |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp      |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp   |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJ_F.cpp          |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJ_VF.cpp         |   6 +-
 .../kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp      |   6 +-
 .../kernel_ElecEw_VdwLJCombGeom_F.cpp         |   6 +-
 .../kernel_ElecEw_VdwLJCombGeom_VF.cpp        |   6 +-
 .../kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp     |   6 +-
 .../kernel_ElecEw_VdwLJCombLB_F.cpp           |   6 +-
 .../kernel_ElecEw_VdwLJCombLB_VF.cpp          |   6 +-
 .../kernel_ElecEw_VdwLJCombLB_VgrpF.cpp       |   6 +-
 .../kernel_ElecEw_VdwLJEwCombGeom_F.cpp       |   6 +-
 .../kernel_ElecEw_VdwLJEwCombGeom_VF.cpp      |   6 +-
 .../kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp   |   6 +-
 .../kernel_ElecEw_VdwLJFSw_F.cpp              |   6 +-
 .../kernel_ElecEw_VdwLJFSw_VF.cpp             |   6 +-
 .../kernel_ElecEw_VdwLJFSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecEw_VdwLJPSw_F.cpp              |   6 +-
 .../kernel_ElecEw_VdwLJPSw_VF.cpp             |   6 +-
 .../kernel_ElecEw_VdwLJPSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecEw_VdwLJ_F.cpp                 |   6 +-
 .../kernel_ElecEw_VdwLJ_VF.cpp                |   6 +-
 .../kernel_ElecEw_VdwLJ_VgrpF.cpp             |   6 +-
 ...ernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp |   6 +-
 ...rnel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp |   6 +-
 ...l_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp |   6 +-
 ...kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp |   6 +-
 ...nel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp |   6 +-
 ...nel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp |   6 +-
 ...el_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp |   6 +-
 ...ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp    |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp   |   6 +-
 ...kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp    |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp   |   6 +-
 ...kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJ_F.cpp       |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp      |   6 +-
 .../kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp   |   6 +-
 .../kernel_ElecQSTab_VdwLJCombGeom_F.cpp      |   6 +-
 .../kernel_ElecQSTab_VdwLJCombGeom_VF.cpp     |   6 +-
 .../kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp  |   6 +-
 .../kernel_ElecQSTab_VdwLJCombLB_F.cpp        |   6 +-
 .../kernel_ElecQSTab_VdwLJCombLB_VF.cpp       |   6 +-
 .../kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp    |   6 +-
 .../kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp    |   6 +-
 .../kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp   |   6 +-
 ...kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp |   6 +-
 .../kernel_ElecQSTab_VdwLJFSw_F.cpp           |   6 +-
 .../kernel_ElecQSTab_VdwLJFSw_VF.cpp          |   6 +-
 .../kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp       |   6 +-
 .../kernel_ElecQSTab_VdwLJPSw_F.cpp           |   6 +-
 .../kernel_ElecQSTab_VdwLJPSw_VF.cpp          |   6 +-
 .../kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp       |   6 +-
 .../kernel_ElecQSTab_VdwLJ_F.cpp              |   6 +-
 .../kernel_ElecQSTab_VdwLJ_VF.cpp             |   6 +-
 .../kernel_ElecQSTab_VdwLJ_VgrpF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJCombGeom_F.cpp         |   6 +-
 .../kernel_ElecRF_VdwLJCombGeom_VF.cpp        |   6 +-
 .../kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp     |   6 +-
 .../kernel_ElecRF_VdwLJCombLB_F.cpp           |   6 +-
 .../kernel_ElecRF_VdwLJCombLB_VF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJCombLB_VgrpF.cpp       |   6 +-
 .../kernel_ElecRF_VdwLJEwCombGeom_F.cpp       |   6 +-
 .../kernel_ElecRF_VdwLJEwCombGeom_VF.cpp      |   6 +-
 .../kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp   |   6 +-
 .../kernel_ElecRF_VdwLJFSw_F.cpp              |   6 +-
 .../kernel_ElecRF_VdwLJFSw_VF.cpp             |   6 +-
 .../kernel_ElecRF_VdwLJFSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJPSw_F.cpp              |   6 +-
 .../kernel_ElecRF_VdwLJPSw_VF.cpp             |   6 +-
 .../kernel_ElecRF_VdwLJPSw_VgrpF.cpp          |   6 +-
 .../kernel_ElecRF_VdwLJ_F.cpp                 |   6 +-
 .../kernel_ElecRF_VdwLJ_VF.cpp                |   6 +-
 .../kernel_ElecRF_VdwLJ_VgrpF.cpp             |   6 +-
 .../nbnxm/kernels_simd_4xm/kernel_outer.h     |  12 +-
 .../nbnxm/kernels_simd_4xm/kernel_prune.cpp   |   2 +-
 .../nbnxm/kernels_simd_4xm/kernel_prune.h     |   1 +
 src/gromacs/nbnxm/kernels_simd_4xm/kernels.h  | 120 +++++++++---------
 src/gromacs/nbnxm/nbnxm.cpp                   |   2 +-
 src/gromacs/nbnxm/nbnxm.h                     |  14 +-
 src/gromacs/nbnxm/nbnxm_geometry.cpp          |   3 +-
 src/gromacs/nbnxm/nbnxm_geometry.h            |   4 +-
 src/gromacs/nbnxm/nbnxm_gpu.h                 |   2 +-
 src/gromacs/nbnxm/nbnxm_setup.cpp             |   8 +-
 src/gromacs/nbnxm/opencl/nbnxm_ocl_consts.h   |   2 +
 .../nbnxm/opencl/nbnxm_ocl_data_mgmt.cpp      |   1 +
 src/gromacs/nbnxm/pairlist.cpp                |  10 +-
 src/gromacs/nbnxm/pairlist.h                  |   3 +-
 src/gromacs/nbnxm/pairlist_tuning.cpp         |   4 +-
 src/gromacs/nbnxm/pairlist_tuning.h           |   2 +-
 src/gromacs/nbnxm/pairlistparams.cpp          |   3 +-
 src/gromacs/nbnxm/pairlistset.h               |   2 +-
 src/gromacs/nbnxm/pairlistwork.h              |   4 +-
 src/gromacs/nbnxm/pairsearch.cpp              |   3 +-
 src/gromacs/nbnxm/pairsearch.h                |   4 +-
 232 files changed, 871 insertions(+), 810 deletions(-)
 create mode 100644 docs/doxygen/lib/nbnxm.md

diff --git a/docs/doxygen/lib/nbnxm.md b/docs/doxygen/lib/nbnxm.md
new file mode 100644
index 0000000000..c0c33386a0
--- /dev/null
+++ b/docs/doxygen/lib/nbnxm.md
@@ -0,0 +1,30 @@
+The NxM atom-cluster non-bonded algorithm {#page_nbnxm}
+=========================================
+
+The algorithm
+=============
+
+Computing non-bonded pair interactions is the most time consuming part
+of most molecular dynamics simulations. It is therefore necessary to
+(highly) optimize this to achieve good simulation performance.
+The standard atom pair lists are not a good match for modern SIMD
+(single-instruction multiple-data) CPU architectures, nor for GPGPU
+accelerators. To achieve higher (cache) data reuse and instruction
+parallelism, we cluster atoms in groups of size N and M, where N and M
+are either the same or differ by a factor of 2. This is done by spatial
+gridding and binning. We then construct a pair list between these
+clusters instead of single atoms. This not only leads to a smaller list,
+but also regularizes the data access. The non-bonded pair-interaction
+kernels can then compute interactions between all atom-pairs in
+a cluster-pair simulatenously. For GPUs there is another layer:
+superclusters consisting of multiple clusters to increase data reuse.
+
+Architecture support
+====================
+
+Currently the module supports 5 different kernel architectures:
+* Plain C++: slow, only for reference.
+* SIMD 4xM: targets CPUs using SIMD intrinsics with N=4 and M=2, 4 or 8, SIMD width 2, 4 or 8.
+* SIMD 2xMM: targets CPUs using SIMD intrinsics with N=4 and M=4 or 8, SIMD width 8 or 16.
+* CUDA: targets CUDA GPUs
+* OpenCL: targets GPUs using OpenCL
diff --git a/docs/doxygen/user/mainpage.md b/docs/doxygen/user/mainpage.md
index 959f1e4176..450d853bd7 100644
--- a/docs/doxygen/user/mainpage.md
+++ b/docs/doxygen/user/mainpage.md
@@ -65,6 +65,8 @@ give an overview of some of the topics that are documented:
  - \subpage page_simd <br/>
    Documentation about the new SIMD module that makes it possible to write
    highly accelerated CPU code that is still portable.
+ - \subpage page_nbnxm <br/>
+   Documentation about the non-bonded module which uses NxM atom clusters.
  - \subpage page_awh <br/>
    Documentation about the accelerated weight histogram (AWH) method,
    which is used for accelerating sampling along reaction coordinates.
diff --git a/src/gromacs/mdlib/sim_util.cpp b/src/gromacs/mdlib/sim_util.cpp
index b9663c8ea5..c34b555985 100644
--- a/src/gromacs/mdlib/sim_util.cpp
+++ b/src/gromacs/mdlib/sim_util.cpp
@@ -87,7 +87,6 @@
 #include "gromacs/mdtypes/simulation_workload.h"
 #include "gromacs/mdtypes/state.h"
 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
-#include "gromacs/nbnxm/atomdata.h"
 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 #include "gromacs/nbnxm/nbnxm.h"
 #include "gromacs/pbcutil/ishift.h"
diff --git a/src/gromacs/nbnxm/atomdata.cpp b/src/gromacs/nbnxm/atomdata.cpp
index 5df9633e27..a4e7b0285d 100644
--- a/src/gromacs/nbnxm/atomdata.cpp
+++ b/src/gromacs/nbnxm/atomdata.cpp
@@ -53,8 +53,6 @@
 #include "gromacs/mdtypes/forcerec.h" // only for GET_CGINFO_*
 #include "gromacs/mdtypes/mdatom.h"
 #include "gromacs/nbnxm/nbnxm.h"
-#include "gromacs/nbnxm/nbnxm_geometry.h"
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/pbcutil/ishift.h"
 #include "gromacs/simd/simd.h"
 #include "gromacs/utility/exceptions.h"
@@ -67,6 +65,8 @@
 
 #include "grid.h"
 #include "gridset.h"
+#include "nbnxm_geometry.h"
+#include "pairlist.h"
 
 using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
 
diff --git a/src/gromacs/nbnxm/clusterdistancekerneltype.h b/src/gromacs/nbnxm/clusterdistancekerneltype.h
index ad5f25f373..0726ab1ac6 100644
--- a/src/gromacs/nbnxm/clusterdistancekerneltype.h
+++ b/src/gromacs/nbnxm/clusterdistancekerneltype.h
@@ -45,10 +45,10 @@
 #ifndef GMX_NBNXM_CLUSTERDISTANCEKERNELTYPE_H
 #define GMX_NBNXM_CLUSTERDISTANCEKERNELTYPE_H
 
-#include "gromacs/nbnxm/atomdata.h"
 #include "gromacs/simd/simd.h"
 #include "gromacs/utility/gmxassert.h"
 
+#include "atomdata.h"
 #include "pairlistparams.h"
 
 //! The types of kernel for calculating the distance between pairs of atom clusters
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu b/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu
index dba1ca8c82..ca88b5e94f 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda.cu
@@ -64,11 +64,11 @@
 #include "gromacs/nbnxm/grid.h"
 #include "gromacs/nbnxm/nbnxm.h"
 #include "gromacs/nbnxm/pairlist.h"
-#include "gromacs/nbnxm/cuda/nbnxm_buffer_ops_kernels.cuh"
 #include "gromacs/timing/gpu_timing.h"
 #include "gromacs/utility/cstringutil.h"
 #include "gromacs/utility/gmxassert.h"
 
+#include "nbnxm_buffer_ops_kernels.cuh"
 #include "nbnxm_cuda_types.h"
 
 /***** The kernel declarations/definitions come here *****/
@@ -82,33 +82,33 @@
  */
 #define FUNCTION_DECLARATION_ONLY
 /** Force only **/
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 /** Force & energy **/
 #define CALC_ENERGIES
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef CALC_ENERGIES
 
 /*** Pair-list pruning kernels ***/
 /** Force only **/
 #define PRUNE_NBL
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 /** Force & energy **/
 #define CALC_ENERGIES
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef CALC_ENERGIES
 #undef PRUNE_NBL
 
 /* Prune-only kernels */
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cuh"
+#include "nbnxm_cuda_kernel_pruneonly.cuh"
 #undef FUNCTION_DECLARATION_ONLY
 
 /* Now generate the function definitions if we are using a single compilation unit. */
 #if GMX_CUDA_NB_SINGLE_COMPILATION_UNIT
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_noprune.cu"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_prune.cu"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_noprune.cu"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_prune.cu"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cu"
+#include "nbnxm_cuda_kernel_F_noprune.cu"
+#include "nbnxm_cuda_kernel_F_prune.cu"
+#include "nbnxm_cuda_kernel_VF_noprune.cu"
+#include "nbnxm_cuda_kernel_VF_prune.cu"
+#include "nbnxm_cuda_kernel_pruneonly.cu"
 #endif /* GMX_CUDA_NB_SINGLE_COMPILATION_UNIT */
 
 namespace Nbnxm
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda.h b/src/gromacs/nbnxm/cuda/nbnxm_cuda.h
index 42296f62b2..f222cddaf0 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda.h
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda.h
@@ -35,6 +35,8 @@
 /*! \internal \file
  * \brief
  * Declares nbnxn cuda cache and texture helper functions
+ *
+ * \ingroup module_nbnxm
  */
 #ifndef GMX_NBNXN_CUDA_NBNXN_CUDA_H
 #define GMX_NBNXN_CUDA_NBNXN_CUDA_H
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_noprune.cu b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_noprune.cu
index 1e17c44a05..6588bae729 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_noprune.cu
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_noprune.cu
@@ -35,8 +35,8 @@
 #include "gmxpre.h"
 
 #include "gromacs/gpu_utils/cudautils.cuh"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_utils.cuh"
 
+#include "nbnxm_cuda_kernel_utils.cuh"
 #include "nbnxm_cuda_types.h"
 
 /* Top-level kernel generation: will generate through multiple
@@ -44,6 +44,6 @@
  * force-only output without pair list pruning;
  */
 #define FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_prune.cu b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_prune.cu
index dde88087ff..ae3a2159e5 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_prune.cu
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_F_prune.cu
@@ -35,8 +35,8 @@
 #include "gmxpre.h"
 
 #include "gromacs/gpu_utils/cudautils.cuh"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_utils.cuh"
 
+#include "nbnxm_cuda_kernel_utils.cuh"
 #include "nbnxm_cuda_types.h"
 
 /* Top-level kernel generation: will generate through multiple
@@ -45,7 +45,7 @@
  */
 #define PRUNE_NBL
 #define FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef PRUNE_NBL
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_noprune.cu b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_noprune.cu
index c6dd13c49a..919cae365f 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_noprune.cu
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_noprune.cu
@@ -35,8 +35,8 @@
 #include "gmxpre.h"
 
 #include "gromacs/gpu_utils/cudautils.cuh"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_utils.cuh"
 
+#include "nbnxm_cuda_kernel_utils.cuh"
 #include "nbnxm_cuda_types.h"
 
 /* Top-level kernel generation: will generate through multiple
@@ -45,7 +45,7 @@
  */
 #define CALC_ENERGIES
 #define FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef CALC_ENERGIES
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_prune.cu b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_prune.cu
index 20e8fedfa0..fda040046d 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_prune.cu
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_VF_prune.cu
@@ -35,8 +35,8 @@
 #include "gmxpre.h"
 
 #include "gromacs/gpu_utils/cudautils.cuh"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_utils.cuh"
 
+#include "nbnxm_cuda_kernel_utils.cuh"
 #include "nbnxm_cuda_types.h"
 
 /* Top-level kernel generation: will generate through multiple
@@ -46,8 +46,8 @@
 #define PRUNE_NBL
 #define CALC_ENERGIES
 #define FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef FUNCTION_DECLARATION_ONLY
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh"
+#include "nbnxm_cuda_kernels.cuh"
 #undef CALC_ENERGIES
 #undef PRUNE_NBL
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cuh b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cuh
index 994951babb..99801a70d1 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cuh
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernel_pruneonly.cuh
@@ -48,9 +48,9 @@
 
 #include "gromacs/gpu_utils/cuda_arch_utils.cuh"
 #include "gromacs/math/utilities.h"
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel_utils.cuh"
 #include "gromacs/pbcutil/ishift.h"
 
+#include "nbnxm_cuda_kernel_utils.cuh"
 #include "nbnxm_cuda_types.h"
 
 /* Note that floating-point constants in CUDA code should be suffixed
diff --git a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh
index 323e6a671f..5bc27a0e98 100644
--- a/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh
+++ b/src/gromacs/nbnxm/cuda/nbnxm_cuda_kernels.cuh
@@ -58,42 +58,42 @@
 
 /* cut-off + V shift LJ */
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecCut_VdwLJ ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w geometric combination rules */
 #define LJ_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecCut_VdwLJCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w LB combination rules */
 #define LJ_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecCut_VdwLJCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w geometric combination rules */
 #define LJ_EWALD_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecCut_VdwLJEwCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w LB combination rules */
 #define LJ_EWALD_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecCut_VdwLJEwCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* F switch LJ */
 #define LJ_FORCE_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecCut_VdwLJFsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_FORCE_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 /* V switch LJ */
 #define LJ_POT_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecCut_VdwLJPsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_POT_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 
@@ -106,42 +106,42 @@
 
 /* cut-off + V shift LJ */
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecRF_VdwLJ ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w geometric combination rules */
 #define LJ_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecRF_VdwLJCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w LB combination rules */
 #define LJ_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecRF_VdwLJCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w geometric combination rules */
 #define LJ_EWALD_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecRF_VdwLJEwCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w LB combination rules */
 #define LJ_EWALD_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecRF_VdwLJEwCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* F switch LJ */
 #define LJ_FORCE_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecRF_VdwLJFsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_FORCE_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 /* V switch LJ */
 #define LJ_POT_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecRF_VdwLJPsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_POT_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 
@@ -154,42 +154,42 @@
 
 /* cut-off + V shift LJ */
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEw_VdwLJ ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w geometric combination rules */
 #define LJ_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEw_VdwLJCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w LB combination rules */
 #define LJ_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEw_VdwLJCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w geometric combination rules */
 #define LJ_EWALD_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEw_VdwLJEwCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w LB combination rules */
 #define LJ_EWALD_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEw_VdwLJEwCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* F switch LJ */
 #define LJ_FORCE_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEw_VdwLJFsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_FORCE_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 /* V switch LJ */
 #define LJ_POT_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEw_VdwLJPsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_POT_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 
@@ -203,42 +203,42 @@
 
 /* cut-off + V shift LJ */
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwTwinCut_VdwLJ ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w geometric combination rules */
 #define LJ_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwTwinCut_VdwLJCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w LB combination rules */
 #define LJ_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwTwinCut_VdwLJCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w geometric combination rules */
 #define LJ_EWALD_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwTwinCut_VdwLJEwCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w LB combination rules */
 #define LJ_EWALD_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwTwinCut_VdwLJEwCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* F switch LJ */
 #define LJ_FORCE_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwTwinCut_VdwLJFsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_FORCE_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 /* V switch LJ */
 #define LJ_POT_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwTwinCut_VdwLJPsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_POT_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 
@@ -251,42 +251,42 @@
 
 /* cut-off + V shift LJ */
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTab_VdwLJ ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w geometric combination rules */
 #define LJ_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTab_VdwLJCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w LB combination rules */
 #define LJ_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTab_VdwLJCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w geometric combination rules */
 #define LJ_EWALD_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTab_VdwLJEwCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w LB combination rules */
 #define LJ_EWALD_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTab_VdwLJEwCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* F switch LJ */
 #define LJ_FORCE_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTab_VdwLJFsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_FORCE_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 /* V switch LJ */
 #define LJ_POT_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTab_VdwLJPsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_POT_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 
@@ -299,42 +299,42 @@
 
 /* cut-off + V shift LJ */
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTabTwinCut_VdwLJ ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w geometric combination rules */
 #define LJ_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTabTwinCut_VdwLJCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* cut-off + V shift LJ w LB combination rules */
 #define LJ_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTabTwinCut_VdwLJCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w geometric combination rules */
 #define LJ_EWALD_COMB_GEOM
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTabTwinCut_VdwLJEwCombGeom ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_GEOM
 #undef NB_KERNEL_FUNC_NAME
 /* LJ-Ewald w LB combination rules */
 #define LJ_EWALD_COMB_LB
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTabTwinCut_VdwLJEwCombLB ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_EWALD_COMB_LB
 #undef NB_KERNEL_FUNC_NAME
 /* F switch LJ */
 #define LJ_FORCE_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTabTwinCut_VdwLJFsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_FORCE_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 /* V switch LJ */
 #define LJ_POT_SWITCH
 #define NB_KERNEL_FUNC_NAME(x, ...) x ## _ElecEwQSTabTwinCut_VdwLJPsw ## __VA_ARGS__
-#include "gromacs/nbnxm/cuda/nbnxm_cuda_kernel.cuh"
+#include "nbnxm_cuda_kernel.cuh"
 #undef LJ_POT_SWITCH
 #undef NB_KERNEL_FUNC_NAME
 
diff --git a/src/gromacs/nbnxm/gpu_jit_support.h b/src/gromacs/nbnxm/gpu_jit_support.h
index 29234e3d42..8aca815097 100644
--- a/src/gromacs/nbnxm/gpu_jit_support.h
+++ b/src/gromacs/nbnxm/gpu_jit_support.h
@@ -37,6 +37,8 @@
  *
  *  \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
  *  \author Mark Abraham <mark.j.abraham@gmail.com>
+ *
+ * \ingroup module_nbnxm
  */
 
 #ifndef GMX_NBNXM_GPU_JIT_SUPPORT_H
diff --git a/src/gromacs/nbnxm/gpu_types_common.h b/src/gromacs/nbnxm/gpu_types_common.h
index 4c57f22bcd..35270a8345 100644
--- a/src/gromacs/nbnxm/gpu_types_common.h
+++ b/src/gromacs/nbnxm/gpu_types_common.h
@@ -44,10 +44,10 @@
 
 #include "config.h"
 
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/utility/enumerationhelpers.h"
 
 #include "locality.h"
+#include "pairlist.h"
 
 #if GMX_GPU == GMX_GPU_OPENCL
 #include "gromacs/gpu_utils/gpuregiontimer_ocl.h"
diff --git a/src/gromacs/nbnxm/grid.cpp b/src/gromacs/nbnxm/grid.cpp
index 54b10eaf28..c6c000bc7c 100644
--- a/src/gromacs/nbnxm/grid.cpp
+++ b/src/gromacs/nbnxm/grid.cpp
@@ -56,13 +56,13 @@
 #include "gromacs/mdlib/gmx_omp_nthreads.h"
 #include "gromacs/mdlib/updategroupscog.h"
 #include "gromacs/mdtypes/forcerec.h" // only for GET_CGINFO_*
-#include "gromacs/nbnxm/atomdata.h"
-#include "gromacs/nbnxm/nbnxm_geometry.h"
 #include "gromacs/simd/simd.h"
 #include "gromacs/simd/vector_operations.h"
 
+#include "atomdata.h"
 #include "boundingboxes.h"
 #include "gridsetdata.h"
+#include "nbnxm_geometry.h"
 #include "pairlistparams.h"
 
 namespace Nbnxm
diff --git a/src/gromacs/nbnxm/gridset.cpp b/src/gromacs/nbnxm/gridset.cpp
index e70763770a..b4d76234bf 100644
--- a/src/gromacs/nbnxm/gridset.cpp
+++ b/src/gromacs/nbnxm/gridset.cpp
@@ -48,9 +48,10 @@
 
 #include "gromacs/mdlib/gmx_omp_nthreads.h"
 #include "gromacs/mdlib/updategroupscog.h"
-#include "gromacs/nbnxm/atomdata.h"
 #include "gromacs/utility/fatalerror.h"
 
+#include "atomdata.h"
+
 namespace Nbnxm
 {
 
diff --git a/src/gromacs/nbnxm/kernel_common.cpp b/src/gromacs/nbnxm/kernel_common.cpp
index 55a470b08c..9c0dc97d8d 100644
--- a/src/gromacs/nbnxm/kernel_common.cpp
+++ b/src/gromacs/nbnxm/kernel_common.cpp
@@ -39,6 +39,7 @@
  * Implements utility functions used by all nbnxm CPU kernels.
  *
  * \author Berk Hess <hess@kth.se>
+ * \ingroup module_nbnxm
  */
 
 #include "gmxpre.h"
diff --git a/src/gromacs/nbnxm/kernel_common.h b/src/gromacs/nbnxm/kernel_common.h
index 7a2b97a7aa..16b3890a40 100644
--- a/src/gromacs/nbnxm/kernel_common.h
+++ b/src/gromacs/nbnxm/kernel_common.h
@@ -39,6 +39,7 @@
  * Declares the nbnxm pair interaction kernel function types and kind counts, also declares utility functions used in nbnxm_kernel.cpp.
  *
  * \author Berk Hess <hess@kth.se>
+ * \ingroup module_nbnxm
  */
 
 #ifndef GMX_NBXNM_KERNEL_COMMON_H
@@ -46,10 +47,11 @@
 
 #include "gromacs/math/vectypes.h"
 /* nbnxn_atomdata_t and nbnxn_pairlist_t could be forward declared, but that requires modifications in all SIMD kernel files */
-#include "gromacs/nbnxm/atomdata.h"
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/utility/real.h"
 
+#include "atomdata.h"
+#include "pairlist.h"
+
 struct interaction_const_t;
 
 // TODO: Consider using one nbk_func type now ener and noener are identical
diff --git a/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_2xmm_kernel.cpp.pre b/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_2xmm_kernel.cpp.pre
index 28a9d829c9..ee2e981a0a 100644
--- a/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_2xmm_kernel.cpp.pre
+++ b/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_2xmm_kernel.cpp.pre
@@ -51,7 +51,7 @@
 {3}
 
 #ifdef {0}
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* {0} */
 
 #ifdef CALC_ENERGIES
@@ -70,7 +70,7 @@ void
 {6}nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef {0}
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* {0} */
 {{
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_4xm_kernel.cpp.pre b/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_4xm_kernel.cpp.pre
index ac936a16a5..ee2e981a0a 100644
--- a/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_4xm_kernel.cpp.pre
+++ b/src/gromacs/nbnxm/kernel_file_generator/kernel_simd_4xm_kernel.cpp.pre
@@ -51,7 +51,7 @@
 {3}
 
 #ifdef {0}
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* {0} */
 
 #ifdef CALC_ENERGIES
@@ -70,7 +70,7 @@ void
 {6}nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef {0}
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* {0} */
 {{
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernel_file_generator/make_verlet_simd_kernel_files.py b/src/gromacs/nbnxm/kernel_file_generator/make_verlet_simd_kernel_files.py
index 1a1ffab795..0f435c2b89 100755
--- a/src/gromacs/nbnxm/kernel_file_generator/make_verlet_simd_kernel_files.py
+++ b/src/gromacs/nbnxm/kernel_file_generator/make_verlet_simd_kernel_files.py
@@ -160,7 +160,7 @@ for type in VerletKernelTypeDict:
     KernelsName = "{0}_simd_{1}".format(KernelNamePrefix,type)
     KernelsFileName = "{0}_simd_{1}".format(KernelFileNamePrefix,type)
     KernelsHeaderFileName = "kernels.h"
-    KernelsHeaderPathName = "gromacs/nbnxm/kernels_simd_{0}/{1}".format(type,KernelsHeaderFileName)
+    KernelsHeaderPathName = "{1}".format(type,KernelsHeaderFileName)
     KernelFunctionLookupTable = {}
     KernelDeclarations = ''
     KernelTemplate = read_kernel_template("{0}_kernel.cpp.pre".format(KernelsFileName))
diff --git a/src/gromacs/nbnxm/kerneldispatch.cpp b/src/gromacs/nbnxm/kerneldispatch.cpp
index f108413439..d7c68273e0 100644
--- a/src/gromacs/nbnxm/kerneldispatch.cpp
+++ b/src/gromacs/nbnxm/kerneldispatch.cpp
@@ -52,23 +52,23 @@
 #include "gromacs/mdtypes/simulation_workload.h"
 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 #include "gromacs/nbnxm/nbnxm.h"
-#include "gromacs/nbnxm/nbnxm_simd.h"
-#include "gromacs/nbnxm/kernels_reference/kernel_gpu_ref.h"
 #include "gromacs/simd/simd.h"
 #include "gromacs/timing/wallcycle.h"
 #include "gromacs/utility/gmxassert.h"
 #include "gromacs/utility/real.h"
 
 #include "kernel_common.h"
+#include "nbnxm_simd.h"
 #include "pairlistset.h"
 #include "pairlistsets.h"
+#include "kernels_reference/kernel_gpu_ref.h"
 #define INCLUDE_KERNELFUNCTION_TABLES
-#include "gromacs/nbnxm/kernels_reference/kernel_ref.h"
+#include "kernels_reference/kernel_ref.h"
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels_simd_2xmm/kernels.h"
 #endif
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels_simd_4xm/kernels.h"
 #endif
 #undef INCLUDE_FUNCTION_TABLES
 
diff --git a/src/gromacs/nbnxm/kernels_reference/kernel_ref.cpp b/src/gromacs/nbnxm/kernels_reference/kernel_ref.cpp
index e953f51e9e..2ae8cb7b90 100644
--- a/src/gromacs/nbnxm/kernels_reference/kernel_ref.cpp
+++ b/src/gromacs/nbnxm/kernels_reference/kernel_ref.cpp
@@ -54,21 +54,21 @@
 /* Analytical reaction-field kernels */
 #define CALC_COUL_RF
 #define LJ_CUT
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_CUT
 #define LJ_FORCE_SWITCH
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_FORCE_SWITCH
 #define LJ_POT_SWITCH
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_POT_SWITCH
 #define LJ_EWALD
 #define LJ_CUT
 #define LJ_EWALD_COMB_GEOM
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_EWALD_COMB_GEOM
 #define LJ_EWALD_COMB_LB
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_EWALD_COMB_LB
 #undef LJ_CUT
 #undef LJ_EWALD
@@ -78,42 +78,42 @@
 /* Tabulated exclusion interaction electrostatics kernels */
 #define CALC_COUL_TAB
 #define LJ_CUT
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_CUT
 #define LJ_FORCE_SWITCH
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_FORCE_SWITCH
 #define LJ_POT_SWITCH
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_POT_SWITCH
 #define LJ_EWALD
 #define LJ_CUT
 #define LJ_EWALD_COMB_GEOM
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_EWALD_COMB_GEOM
 #define LJ_EWALD_COMB_LB
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_EWALD_COMB_LB
 #undef LJ_CUT
 #undef LJ_EWALD
 /* Twin-range cut-off kernels */
 #define VDW_CUTOFF_CHECK
 #define LJ_CUT
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_CUT
 #define LJ_FORCE_SWITCH
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_FORCE_SWITCH
 #define LJ_POT_SWITCH
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_POT_SWITCH
 #define LJ_EWALD
 #define LJ_CUT
 #define LJ_EWALD_COMB_GEOM
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_EWALD_COMB_GEOM
 #define LJ_EWALD_COMB_LB
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_includes.h"
+#include "kernel_ref_includes.h"
 #undef LJ_EWALD_COMB_LB
 #undef LJ_CUT
 #undef LJ_EWALD
diff --git a/src/gromacs/nbnxm/kernels_reference/kernel_ref_includes.h b/src/gromacs/nbnxm/kernels_reference/kernel_ref_includes.h
index e85d399a6c..8b825baf96 100644
--- a/src/gromacs/nbnxm/kernels_reference/kernel_ref_includes.h
+++ b/src/gromacs/nbnxm/kernels_reference/kernel_ref_includes.h
@@ -39,16 +39,16 @@
  */
 
 /* Include the force only kernels */
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_outer.h"
+#include "kernel_ref_outer.h"
 
 /* Include the force+energy kernels */
 #define CALC_ENERGIES
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_outer.h"
+#include "kernel_ref_outer.h"
 #undef CALC_ENERGIES
 
 /* Include the force+energygroups kernels */
 #define CALC_ENERGIES
 #define ENERGY_GROUPS
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_outer.h"
+#include "kernel_ref_outer.h"
 #undef ENERGY_GROUPS
 #undef CALC_ENERGIES
diff --git a/src/gromacs/nbnxm/kernels_reference/kernel_ref_outer.h b/src/gromacs/nbnxm/kernels_reference/kernel_ref_outer.h
index b7eae847b2..c72d85f2ac 100644
--- a/src/gromacs/nbnxm/kernels_reference/kernel_ref_outer.h
+++ b/src/gromacs/nbnxm/kernels_reference/kernel_ref_outer.h
@@ -327,19 +327,19 @@ NBK_FUNC_NAME(_VgrpF) // NOLINT(misc-definitions-in-headers)
             {
 #define CALC_COULOMB
 #define HALF_LJ
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_inner.h"
+#include "kernel_ref_inner.h"
 #undef HALF_LJ
 #undef CALC_COULOMB
             }
             else if (do_coul)
             {
 #define CALC_COULOMB
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_inner.h"
+#include "kernel_ref_inner.h"
 #undef CALC_COULOMB
             }
             else
             {
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_inner.h"
+#include "kernel_ref_inner.h"
             }
 #undef CHECK_EXCLS
             cjind++;
@@ -351,19 +351,19 @@ NBK_FUNC_NAME(_VgrpF) // NOLINT(misc-definitions-in-headers)
             {
 #define CALC_COULOMB
 #define HALF_LJ
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_inner.h"
+#include "kernel_ref_inner.h"
 #undef HALF_LJ
 #undef CALC_COULOMB
             }
             else if (do_coul)
             {
 #define CALC_COULOMB
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_inner.h"
+#include "kernel_ref_inner.h"
 #undef CALC_COULOMB
             }
             else
             {
-#include "gromacs/nbnxm/kernels_reference/kernel_ref_inner.h"
+#include "kernel_ref_inner.h"
             }
         }
 
diff --git a/src/gromacs/nbnxm/kernels_reference/kernel_ref_prune.h b/src/gromacs/nbnxm/kernels_reference/kernel_ref_prune.h
index 03dafde9dc..99ed0aad25 100644
--- a/src/gromacs/nbnxm/kernels_reference/kernel_ref_prune.h
+++ b/src/gromacs/nbnxm/kernels_reference/kernel_ref_prune.h
@@ -39,6 +39,7 @@
  * Declares the C reference pruning only kernel.
  *
  * \author Berk Hess <hess@kth.se>
+ * \ingroup module_nbnxm
  */
 
 #include "gromacs/math/vectypes.h"
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp
index 2c12b9613d..5c31698d4e 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_un
                                                 nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp
index a0d5e95805..099bca66f7 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp
index 0992f32c7f..3803ed0b5d 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu    gm
                                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp
index b9541164ee..b1d19a5773 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_F_2xmm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp
index 86f08d9ea6..e2519030b8 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp
index 686ae1740e..7a31512b03 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp
index 46c1cb41f5..ab8bf8040d 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp
index c4ea31e455..7fa0320f4d 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx
                                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp
index 0431f8cabe..88c5f39e56 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -60,7 +60,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -79,7 +79,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu
                                                       nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp
index 89b648ca12..ed6dafc43c 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp
index 0c69f94c5f..2a7917b77d 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp
index 5ac6d6ce4f..bfb086f41e 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp
index 514d4572a5..c0dffe53e4 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp
index e0d70ae7c3..657b03471f 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp
index bb503f4011..b1044f6ec5 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_F.cpp
index 7a24172c87..6b3fd93397 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJ_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp
index bdc7c9388a..da9c054153 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJ_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *n
                                          nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp
index 2dc0cda73b..50300428a1 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJ_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_F.cpp
index 87d3558ec0..b6323323ef 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_unused *n
                                          nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VF.cpp
index d601be84a4..a97371bd92 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp
index 41be71ad10..c2f21bede4 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_F.cpp
index 76e76cb96e..021a85bcbb 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombLB_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VF.cpp
index 18d5accb97..aa0245727a 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombLB_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp
index f5ca4b17c3..f74e7433ad 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJCombLB_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp
index c39a549480..c3c0816b8c 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJEwCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp
index d86f904289..17958821a7 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp
index 7d20da375a..d590ea066a 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_F.cpp
index b607a344d3..5e14ff7bbe 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJFSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VF.cpp
index 8597f6a5b7..732646f5b4 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJFSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp
index 9cd0fb6960..8e777a8320 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_FORCE_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJFSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_F.cpp
index 34da1266ad..2c320ec859 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJPSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VF.cpp
index 154be9e202..cccdfc3a34 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJPSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp
index 0dfc5fd8a0..bcdb0c1db4 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_POT_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJPSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_F.cpp
index 999bf168fa..6db73c6f5a 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJ_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VF.cpp
index 377ce92e7d..2a74fb38dc 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJ_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VgrpF.cpp
index cd32010336..bac6a86676 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecEw_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJ_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp
index dca9524dec..e72c95c2e0 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx
                                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp
index 0e4c780ca8..bef740b5a8 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF_2xmm(const NbnxnPairlistCpu    gm
                                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp
index cb9637b6b2..64272eafc5 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu
                                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp
index 0c8de6f915..c767d799b1 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_F_2xmm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp
index 168c823712..7fb5f93ef0 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VF_2xmm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp
index 0036b8d7dd..136fd24d4f 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF_2xmm(const NbnxnPairlistCpu    g
                                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp
index cf4ede9ec1..ca251ee586 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F_2xmm(const NbnxnPairlistCpu    g
                                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp
index d9ec2003c6..c26946ed8a 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF_2xmm(const NbnxnPairlistCpu
                                                       nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp
index 39f729ce54..c350fbf0b7 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -60,7 +60,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -79,7 +79,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu
                                                          nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp
index fae093689b..77788e08da 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_F_2xmm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp
index b3579eb145..dccb70c6e3 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp
index 8919a06b50..a956a04f0c 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp
index 3bb989c42e..3c7f67039b 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_F_2xmm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp
index 08bef3cc56..836e7e660b 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp
index 6166ff36cf..4ac50a3c7f 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp
index b4b6d4af71..527889957c 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_F_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp
index 4497b91beb..d2a9400d45 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp
index 75b56bdbaa..43b688df54 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp
index 1ff3a3cfb7..a607976942 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp
index a5e9f33662..b33b4cf338 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp
index eb3e403ec4..3185749bfc 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_un
                                                 nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_F.cpp
index d2fe63aa46..f2258ecbf4 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombLB_F_2xmm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp
index 0bab25b83e..0d22200922 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombLB_VF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp
index 8f08aeeb5c..57f5a21772 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombLB_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp
index da06a4c0d6..50e242022f 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp
index fe549484b1..e38e5ce8dd 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp
index d21472cc81..d69c2350d0 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_F.cpp
index 0732716623..492025318f 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJFSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VF.cpp
index 7d9637b9c2..cb675bfe07 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJFSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp
index 7d676edba0..1202bff8b0 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_FORCE_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJFSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_F.cpp
index 9c00eb704a..7c329247d0 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJPSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VF.cpp
index c010f66185..532ff54c75 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJPSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp
index 23aa6ed378..4d63e2f322 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_POT_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJPSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_F.cpp
index 999cde9ced..d588630056 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJ_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VF.cpp
index 3b23d4784a..df813fd2e7 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJ_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp
index d060c98cef..303e1f529b 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJ_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_F.cpp
index 86d90751d9..a743150246 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_unused *n
                                          nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VF.cpp
index 99818fc5cd..e8f402a7b4 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp
index b387862bfe..68854bb568 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_F.cpp
index 516b2e8afb..703f3dde71 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombLB_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VF.cpp
index ded7cc83b6..dc3c782d68 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombLB_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp
index 5ae8e705bb..d723ea8ace 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJCombLB_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp
index 9d4383db77..c473a78bb0 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJEwCombGeom_F_2xmm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp
index b46b3ac84d..c7d60f5fa3 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VF_2xmm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp
index adbac96605..96e03a2739 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_F.cpp
index ab6ecca7d6..3c2f1c0893 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJFSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VF.cpp
index 3fbc4884f1..8d9b89a773 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJFSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp
index 5bd8c696b2..109fd0d5bf 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_FORCE_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJFSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_F.cpp
index 5fc6411b0b..6ff8191c8b 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJPSw_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VF.cpp
index 09b6103caf..baaf19abca 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJPSw_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp
index 5e5ea68601..a34e228761 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_POT_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJPSw_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_F.cpp
index bbbe4ed220..db7683699e 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJ_F_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VF.cpp
index 64dc8cfc81..3876f74830 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJ_VF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VgrpF.cpp
index 2351f9491c..86f7a27499 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_ElecRF_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_2XNN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJ_VgrpF_2xmm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_2XNN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h
index 5fda72098b..c69a9caea4 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_outer.h
@@ -572,13 +572,13 @@
 #define CHECK_EXCLS
             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
             {
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_inner.h"
+#include "kernel_inner.h"
                 cjind++;
             }
 #undef CHECK_EXCLS
             for (; (cjind < cjind1); cjind++)
             {
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_inner.h"
+#include "kernel_inner.h"
             }
 #undef HALF_LJ
 #undef CALC_COULOMB
@@ -590,13 +590,13 @@
 #define CHECK_EXCLS
             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
             {
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_inner.h"
+#include "kernel_inner.h"
                 cjind++;
             }
 #undef CHECK_EXCLS
             for (; (cjind < cjind1); cjind++)
             {
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_inner.h"
+#include "kernel_inner.h"
             }
 #undef CALC_COULOMB
         }
@@ -606,13 +606,13 @@
 #define CHECK_EXCLS
             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
             {
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_inner.h"
+#include "kernel_inner.h"
                 cjind++;
             }
 #undef CHECK_EXCLS
             for (; (cjind < cjind1); cjind++)
             {
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_inner.h"
+#include "kernel_inner.h"
             }
         }
 #undef CALC_LJ
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.cpp b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.cpp
index 30e67265af..1dc9024565 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.cpp
@@ -44,7 +44,7 @@
 
 #ifdef GMX_NBNXN_SIMD_2XNN
 #define GMX_SIMD_J_UNROLL_SIZE 2
-#include "gromacs/nbnxm/kernels_simd_2xmm/kernel_common.h"
+#include "kernel_common.h"
 #endif
 
 /* Prune a single nbnxn_pairtlist_t entry with distance rlistInner */
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.h b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.h
index 92504cf9b7..a228123285 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.h
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernel_prune.h
@@ -39,6 +39,7 @@
  * Declares the SIMD 2xNN pruning only kernel.
  *
  * \author Berk Hess <hess@kth.se>
+ * \ingroup module_nbnxm
  */
 
 #include "gromacs/math/vectypes.h"
diff --git a/src/gromacs/nbnxm/kernels_simd_2xmm/kernels.h b/src/gromacs/nbnxm/kernels_simd_2xmm/kernels.h
index e287e8d143..52ba3fc0df 100644
--- a/src/gromacs/nbnxm/kernels_simd_2xmm/kernels.h
+++ b/src/gromacs/nbnxm/kernels_simd_2xmm/kernels.h
@@ -42,36 +42,36 @@
 
 /* Declare all the different kernel functions.
  */
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombLB_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJ_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJFSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJPSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombLB_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJ_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJFSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJPSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombLB_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJ_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJFSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJPSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJ_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VgrpF_2xmm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombLB_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJ_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJFSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJPSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJEwCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombLB_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJ_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJFSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJPSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombLB_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJ_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJFSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJPSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJEwCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJ_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_F_2xmm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_2xmm;
 
 nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombGeom_VF_2xmm;
 nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombLB_VF_2xmm;
@@ -104,36 +104,36 @@ nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VF_2xmm;
 nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VF_2xmm;
 nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_2xmm;
 
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombLB_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJ_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJFSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJPSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJEwCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombLB_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJ_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJFSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJPSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombLB_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJ_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJFSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJPSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJEwCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJ_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_F_2xmm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombLB_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJ_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJFSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJPSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombLB_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJ_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJFSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJPSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombLB_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJ_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJFSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJPSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJ_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VgrpF_2xmm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF_2xmm;
 
 
 
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp
index 7c2a6a5232..56f1e44ae7 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp
index 49d7ef9c23..ffc0b34081 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_un
                                                 nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp
index 791b16808c..bec89c4b04 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    gmx
                                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp
index de50e9a147..69d71d51f8 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_F_4xm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp
index d685fb389a..40cb3d8978 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp
index 86844e2cb0..ef64e7e0b0 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF_4xm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp
index 41295eafd8..aefe586cd7 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp
index 838ee07aaf..6e0969e1c6 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp
index b681ad0f56..52604e9c0c 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -60,7 +60,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -79,7 +79,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    g
                                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp
index 804221c267..d9c0de482c 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp
index e3bf4ee259..0521bdba6c 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp
index a3834564ea..bfc9f0eaa8 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp
index 143052e50c..aa607fa69d 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp
index 885eb5db31..33de743c76 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp
index cc002f4fba..5cf3e42984 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_F.cpp
index c68b5f7468..1fa321b188 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJ_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp
index 4c6207f64c..c0239f5e5c 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJ_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp
index 3fbc1dc0e7..a9fbb7bfd6 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEwTwinCut_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEwTwinCut_VdwLJ_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_F.cpp
index 8b76dff1df..f79a8b4ec2 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VF.cpp
index 38cbaa6d12..ca9f25d6e5 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_unused *n
                                          nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp
index acf0a2da3b..ef045e2b37 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_F.cpp
index 10c8cce853..cd2b5b55d8 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombLB_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                       nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VF.cpp
index 7ff8d6a7f1..4592c0888f 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJCombLB_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp
index 6851a41b9f..1c8a545131 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJCombLB_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp
index 7456858ce1..3d8ce55d0e 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJEwCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp
index c30eefeb3b..09f6bf8838 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp
index 93598b1f9d..935e5abf52 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_F.cpp
index 16fcfb27de..1844fe305d 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJFSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VF.cpp
index c28305e3d9..af339d9eb5 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJFSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp
index 6c47ff6139..92e4320554 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_FORCE_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJFSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_F.cpp
index ef6681da4f..73d27bc5e7 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJPSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VF.cpp
index 16618de5da..46d71ab0be 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJPSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp
index bc940bd8f0..a8b471acd2 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_POT_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJPSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_F.cpp
index e3ef804746..d98a3070f3 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJ_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                 nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VF.cpp
index 0d271ec000..f56f84ac16 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecEw_VdwLJ_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VgrpF.cpp
index 3ed03df65c..c678781073 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecEw_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_EWALD
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecEw_VdwLJ_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp
index 5fe162e16f..9da9e06358 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_
                                                   nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp
index 6ab41b9ede..e52981e755 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx
                                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp
index 51a3671ae0..d2096a3a40 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF_4xm(const NbnxnPairlistCpu
                                                       nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp
index eb1a0dfe93..2cd9431c68 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_F_4xm(const NbnxnPairlistCpu    gmx_un
                                                 nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp
index 2f7261c2fa..4b2ec4f101 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VF_4xm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp
index fee9517837..2e72c533d1 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF_4xm(const NbnxnPairlistCpu    gm
                                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp
index b0cacb3b39..4b8ee4d3f6 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F_4xm(const NbnxnPairlistCpu    gm
                                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp
index fac9718044..dc1a1b329e 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VF_4xm(const NbnxnPairlistCpu    g
                                                      nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp
index ee0a23c5fe..995d645030 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -60,7 +60,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -79,7 +79,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF_4xm(const NbnxnPairlistCpu
                                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp
index 6719541a0b..1c208163c6 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_F_4xm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp
index 627807ba89..3cca2781e7 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp
index fa31c98673..dc1532155e 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp
index 25b2560937..5f18a79d02 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_F_4xm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp
index 3429889751..156a9637bf 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp
index 5953b4cbba..0aaab4dcbf 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp
index fb636bb806..4372f4eee6 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_F_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp
index a716118b71..60b4030f2d 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VF_4xm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp
index f758f1c9ec..698900695b 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTabTwinCut_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define VDW_CUTOFF_CHECK /* Use twin-range cut-off */
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp
index 369bdefac6..d6bd1c18db 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp
index 5f000a5ca0..19fac05b23 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp
index d0d81bfc08..6df4d69625 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unu
                                                nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_F.cpp
index 3c7ee01f69..4b2b6b7582 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombLB_F_4xm(const NbnxnPairlistCpu    gmx_unused *n
                                          nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp
index bb0069b7c2..b9310d5381 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombLB_VF_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp
index b1a8763dc9..4946c77b66 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJCombLB_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp
index 1817c15774..6efe19d96d 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_unuse
                                              nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp
index 5ba3421bc2..b117989e3e 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp
index 7f1f2c3370..77a76c69be 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    gmx_u
                                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_F.cpp
index 2db415656d..92f4a0e6fb 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJFSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                       nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VF.cpp
index 9a2c8694c9..ec53a50c73 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJFSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp
index 40f61aafbf..71714182b9 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_FORCE_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJFSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_F.cpp
index 7f1c0178ee..bbd30f5979 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJPSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                       nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VF.cpp
index b53b9a92d6..38aabeb1e4 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJPSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp
index e7e875982a..d116aacc11 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_POT_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJPSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_F.cpp
index c651a628de..87f945393f 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJ_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VF.cpp
index 6a3df8e756..c9a7da6779 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecQSTab_VdwLJ_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp
index 846566ce99..170607e108 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecQSTab_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_TAB
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecQSTab_VdwLJ_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_F.cpp
index b8b18d758b..303c1c7fee 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_unused *nb
                                         nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VF.cpp
index 27bbc7a76a..4cfcdb74df 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_unused *n
                                          nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp
index 85d06fe7e1..6798cc2133 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused
                                             nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_F.cpp
index 683922db9a..15c19aef02 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombLB_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                       nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VF.cpp
index ca04a22f79..7b943bc3d8 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJCombLB_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp
index 740daa964e..0e3204140b 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJCombLB_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJCombLB_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp
index dfbbedd6dd..ceeeb64967 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJEwCombGeom_F_4xm(const NbnxnPairlistCpu    gmx_unused *
                                           nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp
index de0d641f09..d3fc817bba 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VF_4xm(const NbnxnPairlistCpu    gmx_unused
                                            nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp
index fe4c3ae1d8..8bff4e688f 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJEwCombGeom_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -59,7 +59,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -78,7 +78,7 @@ nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unus
                                               nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_F.cpp
index f8f6fb647d..7f81b50a86 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJFSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VF.cpp
index a639ba911a..2d9aa86a0a 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_FORCE_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJFSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp
index 577788998f..1816abf2d9 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJFSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_FORCE_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJFSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_F.cpp
index 3a82d95e0c..f2a83db3b5 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJPSw_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                    nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VF.cpp
index 0e0223a9f1..720144261f 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_POT_SWITCH
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJPSw_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp
index bbfaa18662..0fa843ae5f 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJPSw_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_POT_SWITCH
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJPSw_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl
                                        nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_F.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_F.cpp
index 22a0accf01..f2a47c6c2c 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_F.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_F.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 /* Will not calculate energies */
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJ_F_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                 nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VF.cpp
index f55d76a336..2efd7e5360 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -57,7 +57,7 @@
 #define CALC_ENERGIES
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -76,7 +76,7 @@ nbnxm_kernel_ElecRF_VdwLJ_VF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                  nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VgrpF.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VgrpF.cpp
index 18f89e1c46..954585210b 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VgrpF.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_ElecRF_VdwLJ_VgrpF.cpp
@@ -49,7 +49,7 @@
 #include "gromacs/nbnxm/nbnxm_simd.h"
 
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernels.h"
+#include "kernels.h"
 
 #define CALC_COUL_RF
 #define LJ_CUT
@@ -58,7 +58,7 @@
 #define ENERGY_GROUPS
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif /* GMX_NBNXN_SIMD_4XN */
 
 #ifdef CALC_ENERGIES
@@ -77,7 +77,7 @@ nbnxm_kernel_ElecRF_VdwLJ_VgrpF_4xm(const NbnxnPairlistCpu    gmx_unused *nbl,
                                     nbnxn_atomdata_output_t   gmx_unused *out)
 #endif /* CALC_ENERGIES */
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h"
+#include "kernel_outer.h"
 #else /* GMX_NBNXN_SIMD_4XN */
 {
 /* No need to call gmx_incons() here, because the only function
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h
index 217985d56b..be3eeeba1f 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_outer.h
@@ -616,13 +616,13 @@
 #define CHECK_EXCLS
             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
             {
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_inner.h"
+#include "kernel_inner.h"
                 cjind++;
             }
 #undef CHECK_EXCLS
             for (; (cjind < cjind1); cjind++)
             {
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_inner.h"
+#include "kernel_inner.h"
             }
 #undef HALF_LJ
 #undef CALC_COULOMB
@@ -634,13 +634,13 @@
 #define CHECK_EXCLS
             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
             {
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_inner.h"
+#include "kernel_inner.h"
                 cjind++;
             }
 #undef CHECK_EXCLS
             for (; (cjind < cjind1); cjind++)
             {
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_inner.h"
+#include "kernel_inner.h"
             }
 #undef CALC_COULOMB
         }
@@ -650,13 +650,13 @@
 #define CHECK_EXCLS
             while (cjind < cjind1 && nbl->cj[cjind].excl != NBNXN_INTERACTION_MASK_ALL)
             {
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_inner.h"
+#include "kernel_inner.h"
                 cjind++;
             }
 #undef CHECK_EXCLS
             for (; (cjind < cjind1); cjind++)
             {
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_inner.h"
+#include "kernel_inner.h"
             }
         }
 #undef CALC_LJ
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.cpp b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.cpp
index 6d46acff1e..c162aee5dd 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.cpp
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.cpp
@@ -44,7 +44,7 @@
 
 #ifdef GMX_NBNXN_SIMD_4XN
 #define GMX_SIMD_J_UNROLL_SIZE 1
-#include "gromacs/nbnxm/kernels_simd_4xm/kernel_common.h"
+#include "kernel_common.h"
 #endif
 
 /* Prune a single nbnxn_pairtlist_t entry with distance rlistInner */
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.h b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.h
index 47950fc86b..4e0ad9f956 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.h
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernel_prune.h
@@ -39,6 +39,7 @@
  * Declares the SIMD 4xN pruning only kernel.
  *
  * \author Berk Hess <hess@kth.se>
+ * \ingroup module_nbnxm
  */
 
 #include "gromacs/math/vectypes.h"
diff --git a/src/gromacs/nbnxm/kernels_simd_4xm/kernels.h b/src/gromacs/nbnxm/kernels_simd_4xm/kernels.h
index 3ad2c58784..afe1cff7bc 100644
--- a/src/gromacs/nbnxm/kernels_simd_4xm/kernels.h
+++ b/src/gromacs/nbnxm/kernels_simd_4xm/kernels.h
@@ -42,36 +42,36 @@
 
 /* Declare all the different kernel functions.
  */
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombLB_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJ_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJFSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJPSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombLB_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJ_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJFSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJPSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombLB_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJ_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJFSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJPSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJ_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VgrpF_4xm;
-nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF_4xm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombLB_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJ_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJFSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJPSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJEwCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombLB_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJ_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJFSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJPSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombLB_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJ_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJFSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJPSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJEwCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJ_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_F_4xm;
+nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_4xm;
 
 nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombGeom_VF_4xm;
 nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombLB_VF_4xm;
@@ -104,36 +104,36 @@ nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VF_4xm;
 nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VF_4xm;
 nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VF_4xm;
 
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJCombLB_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJ_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJFSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJPSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecRF_VdwLJEwCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJCombLB_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJ_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJFSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJPSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJCombLB_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJ_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJFSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJPSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEw_VdwLJEwCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJ_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_F_4xm;
-nbk_func_noener       nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_F_4xm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJCombLB_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJ_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJFSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJPSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecRF_VdwLJEwCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJCombLB_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJ_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJFSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJPSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTab_VdwLJEwCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJCombLB_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJ_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJFSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJPSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecQSTabTwinCut_VdwLJEwCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJCombLB_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJ_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJFSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJPSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEw_VdwLJEwCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombGeom_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJCombLB_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJ_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJFSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJPSw_VgrpF_4xm;
+nbk_func_ener         nbnxm_kernel_ElecEwTwinCut_VdwLJEwCombGeom_VgrpF_4xm;
 
 
 
diff --git a/src/gromacs/nbnxm/nbnxm.cpp b/src/gromacs/nbnxm/nbnxm.cpp
index 5c6967bb65..88e6169b71 100644
--- a/src/gromacs/nbnxm/nbnxm.cpp
+++ b/src/gromacs/nbnxm/nbnxm.cpp
@@ -46,9 +46,9 @@
 #include "nbnxm.h"
 
 #include "gromacs/domdec/domdec_struct.h"
-#include "gromacs/nbnxm/atomdata.h"
 #include "gromacs/timing/wallcycle.h"
 
+#include "atomdata.h"
 #include "pairlistsets.h"
 #include "pairsearch.h"
 
diff --git a/src/gromacs/nbnxm/nbnxm.h b/src/gromacs/nbnxm/nbnxm.h
index 991f21692e..3be931f7d9 100644
--- a/src/gromacs/nbnxm/nbnxm.h
+++ b/src/gromacs/nbnxm/nbnxm.h
@@ -80,6 +80,18 @@
  * TODO: add more authors!
  */
 
+/*! \libinternal
+ * \defgroup module_nbnxm Non-bonded pair interactions
+ * \ingroup group_mdrun
+ * \brief
+ * Implements non-bonded pair interaction functionality for NxM atom clusters.
+ *
+ * This module provides methods to, very efficiently, compute non-bonded
+ * pair interactions on CPUs as well as accelerators. It also provides
+ * a method to construct the NxM atom-cluster pair-list required for
+ * computing these non-bonded iteractions.
+ */
+
 /*! \libinternal \file
  *
  * \brief This file contains the public interface of the nbnxm module
@@ -91,7 +103,7 @@
  * \author Szilárd Páll <pall.szilard@gmail.com>
  *
  * \inlibraryapi
- * \ingroup __module_nbnxm
+ * \ingroup module_nbnxm
  */
 
 
diff --git a/src/gromacs/nbnxm/nbnxm_geometry.cpp b/src/gromacs/nbnxm/nbnxm_geometry.cpp
index c164034c79..ff1f6ee1d1 100644
--- a/src/gromacs/nbnxm/nbnxm_geometry.cpp
+++ b/src/gromacs/nbnxm/nbnxm_geometry.cpp
@@ -38,10 +38,11 @@
 #include "nbnxm_geometry.h"
 
 #include "gromacs/nbnxm/nbnxm.h"
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/utility/fatalerror.h"
 #include "gromacs/utility/real.h"
 
+#include "pairlist.h"
+
 /* Clusters at the cut-off only increase rlist by 60% of their size */
 static constexpr real c_nbnxnRlistIncreaseOutsideFactor = 0.6;
 
diff --git a/src/gromacs/nbnxm/nbnxm_geometry.h b/src/gromacs/nbnxm/nbnxm_geometry.h
index c43e22e7ba..5c0e217ffd 100644
--- a/src/gromacs/nbnxm/nbnxm_geometry.h
+++ b/src/gromacs/nbnxm/nbnxm_geometry.h
@@ -38,10 +38,12 @@
 
 #include "gromacs/math/vectypes.h"
 #include "gromacs/nbnxm/nbnxm.h"
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/simd/simd.h"
 #include "gromacs/utility/fatalerror.h"
 
+#include "pairlist.h"
+
+
 /* Returns the base-2 log of n.
  * Generates a fatal error when n is not an integer power of 2.
  */
diff --git a/src/gromacs/nbnxm/nbnxm_gpu.h b/src/gromacs/nbnxm/nbnxm_gpu.h
index f661a19ca1..ae396e9089 100644
--- a/src/gromacs/nbnxm/nbnxm_gpu.h
+++ b/src/gromacs/nbnxm/nbnxm_gpu.h
@@ -45,10 +45,10 @@
 
 #include "gromacs/gpu_utils/gpu_macros.h"
 #include "gromacs/math/vectypes.h"
-#include "gromacs/nbnxm/atomdata.h"
 #include "gromacs/utility/basedefinitions.h"
 #include "gromacs/utility/real.h"
 
+#include "atomdata.h"
 #include "gpu_types.h"
 #include "locality.h"
 
diff --git a/src/gromacs/nbnxm/nbnxm_setup.cpp b/src/gromacs/nbnxm/nbnxm_setup.cpp
index 8c0b736083..cd13b9cc44 100644
--- a/src/gromacs/nbnxm/nbnxm_setup.cpp
+++ b/src/gromacs/nbnxm/nbnxm_setup.cpp
@@ -49,19 +49,19 @@
 #include "gromacs/mdtypes/commrec.h"
 #include "gromacs/mdtypes/forcerec.h"
 #include "gromacs/mdtypes/inputrec.h"
-#include "gromacs/nbnxm/atomdata.h"
 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 #include "gromacs/nbnxm/nbnxm.h"
-#include "gromacs/nbnxm/nbnxm_geometry.h"
-#include "gromacs/nbnxm/nbnxm_simd.h"
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/nbnxm/pairlist_tuning.h"
 #include "gromacs/simd/simd.h"
 #include "gromacs/utility/fatalerror.h"
 #include "gromacs/utility/logger.h"
 
+#include "atomdata.h"
 #include "gpu_types.h"
 #include "grid.h"
+#include "nbnxm_geometry.h"
+#include "nbnxm_simd.h"
+#include "pairlist.h"
 #include "pairlistset.h"
 #include "pairlistsets.h"
 #include "pairsearch.h"
diff --git a/src/gromacs/nbnxm/opencl/nbnxm_ocl_consts.h b/src/gromacs/nbnxm/opencl/nbnxm_ocl_consts.h
index c2380121d4..a49596223f 100644
--- a/src/gromacs/nbnxm/opencl/nbnxm_ocl_consts.h
+++ b/src/gromacs/nbnxm/opencl/nbnxm_ocl_consts.h
@@ -39,6 +39,8 @@
  * \brief Macros defining platform-dependent defaults for the prune kernel's j4 processing concurrency.
  *
  *  The GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY macro allows compile-time override.
+ *
+ * \ingroup module_nbnxm
  */
 /*! @{ */
 #define GMX_NBNXN_PRUNE_KERNEL_J4_CONCURRENCY_DEFAULT       4
diff --git a/src/gromacs/nbnxm/opencl/nbnxm_ocl_data_mgmt.cpp b/src/gromacs/nbnxm/opencl/nbnxm_ocl_data_mgmt.cpp
index 4123a65e39..35d5fea8d9 100644
--- a/src/gromacs/nbnxm/opencl/nbnxm_ocl_data_mgmt.cpp
+++ b/src/gromacs/nbnxm/opencl/nbnxm_ocl_data_mgmt.cpp
@@ -39,6 +39,7 @@
  *  \author Dimitrios Karkoulis <dimitris.karkoulis@gmail.com>
  *  \author Teemu Virolainen <teemu@streamcomputing.eu>
  *  \author Szilárd Páll <pall.szilard@gmail.com>
+ *  \ingroup module_nbnxm
  */
 #include "gmxpre.h"
 
diff --git a/src/gromacs/nbnxm/pairlist.cpp b/src/gromacs/nbnxm/pairlist.cpp
index 8b55a3a9dd..4ab89bc072 100644
--- a/src/gromacs/nbnxm/pairlist.cpp
+++ b/src/gromacs/nbnxm/pairlist.cpp
@@ -53,10 +53,7 @@
 #include "gromacs/mdlib/gmx_omp_nthreads.h"
 #include "gromacs/mdtypes/group.h"
 #include "gromacs/mdtypes/md_enums.h"
-#include "gromacs/nbnxm/atomdata.h"
 #include "gromacs/nbnxm/gpu_data_mgmt.h"
-#include "gromacs/nbnxm/nbnxm_geometry.h"
-#include "gromacs/nbnxm/nbnxm_simd.h"
 #include "gromacs/pbcutil/ishift.h"
 #include "gromacs/pbcutil/pbc.h"
 #include "gromacs/simd/simd.h"
@@ -67,9 +64,12 @@
 #include "gromacs/utility/gmxomp.h"
 #include "gromacs/utility/smalloc.h"
 
+#include "atomdata.h"
 #include "boundingboxes.h"
 #include "clusterdistancekerneltype.h"
 #include "gridset.h"
+#include "nbnxm_geometry.h"
+#include "nbnxm_simd.h"
 #include "pairlistset.h"
 #include "pairlistsets.h"
 #include "pairlistwork.h"
@@ -1079,10 +1079,10 @@ makeClusterListSimple(const Grid               &jGrid,
 }
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "gromacs/nbnxm/pairlist_simd_4xm.h"
+#include "pairlist_simd_4xm.h"
 #endif
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "gromacs/nbnxm/pairlist_simd_2xmm.h"
+#include "pairlist_simd_2xmm.h"
 #endif
 
 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
diff --git a/src/gromacs/nbnxm/pairlist.h b/src/gromacs/nbnxm/pairlist.h
index 19d898fcd0..6e2cf8bf55 100644
--- a/src/gromacs/nbnxm/pairlist.h
+++ b/src/gromacs/nbnxm/pairlist.h
@@ -48,8 +48,7 @@
 
 // This file with constants is separate from this file to be able
 // to include it during OpenCL jitting without including config.h
-#include "gromacs/nbnxm/constants.h"
-
+#include "constants.h"
 #include "locality.h"
 #include "pairlistparams.h"
 
diff --git a/src/gromacs/nbnxm/pairlist_tuning.cpp b/src/gromacs/nbnxm/pairlist_tuning.cpp
index 91b4477e37..d8a8859be8 100644
--- a/src/gromacs/nbnxm/pairlist_tuning.cpp
+++ b/src/gromacs/nbnxm/pairlist_tuning.cpp
@@ -38,7 +38,7 @@
  * \brief Implements functions for tuning adjustable parameters for the nbnxn non-bonded search and interaction kernels
  *
  * \author Berk Hess <hess@kth.se>
- * \ingroup __module_nb_verlet
+ * \ingroup module_nbnxm
  */
 
 #include "gmxpre.h"
@@ -60,7 +60,6 @@
 #include "gromacs/mdtypes/inputrec.h"
 #include "gromacs/mdtypes/interaction_const.h"
 #include "gromacs/mdtypes/state.h"
-#include "gromacs/nbnxm/nbnxm_geometry.h"
 #include "gromacs/pbcutil/pbc.h"
 #include "gromacs/topology/topology.h"
 #include "gromacs/utility/cstringutil.h"
@@ -70,6 +69,7 @@
 #include "gromacs/utility/strconvert.h"
 #include "gromacs/utility/stringutil.h"
 
+#include "nbnxm_geometry.h"
 #include "pairlistsets.h"
 
 /*! \brief Returns if we can (heuristically) change nstlist and rlist
diff --git a/src/gromacs/nbnxm/pairlist_tuning.h b/src/gromacs/nbnxm/pairlist_tuning.h
index e3af9a42fb..bde3474bc3 100644
--- a/src/gromacs/nbnxm/pairlist_tuning.h
+++ b/src/gromacs/nbnxm/pairlist_tuning.h
@@ -40,7 +40,7 @@
  * \author Berk Hess <hess@kth.se>
  *
  * \inlibraryapi
- * \ingroup __module_nb_verlet
+ * \ingroup module_nbnxm
  */
 
 #ifndef NBNXM_PAIRLIST_TUNING_H
diff --git a/src/gromacs/nbnxm/pairlistparams.cpp b/src/gromacs/nbnxm/pairlistparams.cpp
index ca27ef5246..6deada0692 100644
--- a/src/gromacs/nbnxm/pairlistparams.cpp
+++ b/src/gromacs/nbnxm/pairlistparams.cpp
@@ -46,9 +46,10 @@
 #include "pairlistparams.h"
 
 #include "gromacs/nbnxm/nbnxm.h"
-#include "gromacs/nbnxm/nbnxm_geometry.h"
 #include "gromacs/utility/gmxassert.h"
 
+#include "nbnxm_geometry.h"
+
 
 PairlistParams::PairlistParams(const Nbnxm::KernelType kernelType,
                                const bool              haveFep,
diff --git a/src/gromacs/nbnxm/pairlistset.h b/src/gromacs/nbnxm/pairlistset.h
index 6fbe2b1792..2fad5d6f17 100644
--- a/src/gromacs/nbnxm/pairlistset.h
+++ b/src/gromacs/nbnxm/pairlistset.h
@@ -52,11 +52,11 @@
 #include <memory>
 
 #include "gromacs/math/vectypes.h"
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/utility/basedefinitions.h"
 #include "gromacs/utility/real.h"
 
 #include "locality.h"
+#include "pairlist.h"
 
 struct nbnxn_atomdata_t;
 struct PairlistParams;
diff --git a/src/gromacs/nbnxm/pairlistwork.h b/src/gromacs/nbnxm/pairlistwork.h
index 77dc96e7c1..5cd0c01fe4 100644
--- a/src/gromacs/nbnxm/pairlistwork.h
+++ b/src/gromacs/nbnxm/pairlistwork.h
@@ -39,7 +39,7 @@
  *
  * \author Berk Hess <hess@kth.se>
  *
- * \ingroup module_nbnxn
+ * \ingroup module_nbnxm
  */
 
 #ifndef GMX_NBNXM_PAIRLISTWORK_H
@@ -48,11 +48,11 @@
 #include <memory>
 #include <vector>
 
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/simd/simd.h"
 
 #include "boundingboxes.h"
 #include "grid.h"
+#include "pairlist.h"
 
 /* Working data for the actual i-supercell during pair search */
 struct NbnxnPairlistCpuWork
diff --git a/src/gromacs/nbnxm/pairsearch.cpp b/src/gromacs/nbnxm/pairsearch.cpp
index 2d425de5ef..5988ccbf22 100644
--- a/src/gromacs/nbnxm/pairsearch.cpp
+++ b/src/gromacs/nbnxm/pairsearch.cpp
@@ -45,9 +45,10 @@
 
 #include "pairsearch.h"
 
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/utility/smalloc.h"
 
+#include "pairlist.h"
+
 
 void SearchCycleCounting::printCycles(FILE                               *fp,
                                       gmx::ArrayRef<const PairsearchWork> work) const
diff --git a/src/gromacs/nbnxm/pairsearch.h b/src/gromacs/nbnxm/pairsearch.h
index 7bf8f421a3..4258932001 100644
--- a/src/gromacs/nbnxm/pairsearch.h
+++ b/src/gromacs/nbnxm/pairsearch.h
@@ -55,14 +55,14 @@
 
 #include "gromacs/domdec/domdec.h"
 #include "gromacs/math/vectypes.h"
-#include "gromacs/nbnxm/atomdata.h"
-#include "gromacs/nbnxm/pairlist.h"
 #include "gromacs/timing/cyclecounter.h"
 #include "gromacs/utility/alignedallocator.h"
 #include "gromacs/utility/arrayref.h"
 #include "gromacs/utility/real.h"
 
+#include "atomdata.h"
 #include "gridset.h"
+#include "pairlist.h"
 
 struct gmx_domdec_zones_t;
 struct PairsearchWork;
-- 
2.22.0