Performance improvements#

Implemented support for PME long-ranged interactions on GPUs#

A single GPU can now be used to accelerate the computation of the long-ranged PME interactions. This feature provides excellent performance improvements, in particular that only 2-4 CPU cores per GPU will be about as fast as the 2016 version that needed many more CPU cores to balance the GPU. Performance on hardware that had good balance of GPU and CPU also shows minor improvements, and the capacity for hardware with strong GPUs to run effective simulations is now greatly improved.

Currently, the GPU used for PME must be either the same GPU as used for the short-ranged interactions and in the same single rank of the simulation, or any GPU used from a PME-only rank. mdrun -pme gpu now requires that PME runs on a GPU, if supported. All CUDA versions and hardware generations supported by GROMACS can run this code path, including CUDA 9.0 and Volta GPUs. However, not all combinations of features are supported with PME on GPUs - notably FEP calculations are not yet available.

The user guide is updated to reflect the new capabilities, and more documentation will be forthcoming.

Added more SIMD intrinsics support for PME spread and gather#

Achieved speedup on Intel KNL processors of around 11% for PME spread/gather on typical simulation systems.

Added SIMD intrinsics version of simple update#

In the simple case of leap-frog without pressure coupling and with at most one temperature-coupling group, the update of velocities and coordinates is now implemented with SIMD intrinsics for improved simulation rate.

Add SIMD intrinsics version of Urey-Bradley angle kernel#

For steps where energies and shift forces are not required, this kernel improves performance, which can otherwise be rate limiting in GPU-accelerated runs, particularly with CHARMM force fields.

Use OpenMP up to 16 threads with AMD Ryzen when automating run setup#

AMD Ryzen appears to always perform slightly better with OpenMP than MPI, up to using all 16 threads on the 8-core die.

128-bit AVX2 SIMD for AMD Ryzen#

While Ryzen supports 256-bit AVX2, the internal units are organized to execute either a single 256-bit instruction or two 128-bit SIMD instruction per cycle. Since most of our kernels are slightly less efficient for wider SIMD, this improves performance by roughly 10%.

Choose faster nbnxn SIMD kernels on AMD Zen#

On AMD Zen, tabulated Ewald kernels are always faster than analytical. And with AVX2_256 2xNN kernels are faster than 4xN. These faster choices are now made based on CpuInfo at run time.

Refs Issue 2328

Enabled group-scheme SIMD with GMX_SIMD=AVX2_128#

The group-scheme kernels can use AVX instructions from either the AVX_128_FMA and AVX_256 extensions. But hardware that supports the new AVX2_128 extensions also supports AVX_256, so we enable such support for the group-scheme kernels.

Detect AVX-512 FMA units to choose best SIMD#

Recent Intel x86 hardware can have multiple AVX-512 FMA units, and the number of those units and the way their use interacts with the way the CPU chooses its clock speed mean that it can be advantageous to avoid using AVX-512 SIMD support in GROMACS if there is only one such unit. Because there is no way to query the hardware to count the number of such units, we run code at CMake and mdrun time to compare the performance from using such units, and recommend the version that is best. This may mean that building GROMACS on the front-end node of the cluster might not suit the compute nodes, even when they are all from the same generation of Intel’s hardware.

Speed up nbnxn buffer clearing#

Tweaked conditional in the nonbonded GPU kernels#

GPU compilers miss an easy optimization of a loop invariant in the inner-lop conditional. Precomputing part of the conditional together with using bitwise instead of logical and/or improves performance with most compilers by up to 5%.