nbnxm_simd.h File Reference

#include "config.h"
#include "gromacs/simd/simd.h"
#include "gromacs/utility/real.h"
#include "gromacs/utility/vectypes.h"
#include "pairlist.h"
#include "pairlistparams.h"

Include dependency graph for nbnxm_simd.h:

This graph shows which files directly or indirectly include this file:

Description

Defines constants used to know which nbNxM kernel flavours (4xM or 2xMM) can be supported by the SIMD layer in use.

Author

Berk Hess hess@.nosp@m.kth..nosp@m.se

Macros
#define	GMX_HAVE_NBNXM_SIMD_2XMM   0
	The nbnxn SIMD 4xN and 2x(N+N) kernels can be added independently. Currently the 2xNN SIMD kernels only make sense with: 8-way SIMD: 4x4 setup, performance wise only useful on CPUs without FMA or on AMD Zen1 16-way SIMD: 4x8 setup, used in single precision with 512 bit wide SIMD.
#define	GMX_HAVE_NBNXM_SIMD_4XM   0

Enumerations
enum	KernelLayout { KernelLayout::r4xM, KernelLayout::r2xMM }
	The types of nbNxM SIMD kernel layout. More...
enum	KernelLayoutClusterRatio { KernelLayoutClusterRatio::JSizeEqualsISize, KernelLayoutClusterRatio::JSizeIsDoubleISize, KernelLayoutClusterRatio::JSizeIsHalfISize }
	List of supported ratios for j-cluster size versus i-cluster sizes. More...

Functions
static constexpr int	sc_iClusterSize (const KernelLayout kernelLayout)
	The NBNxM i-cluster size in atoms for the given NBNxM kernel layout.

Variables
static constexpr bool	sc_haveNbnxmSimd2xmmKernels = 0
	Whether we have support for NBNxM 2xM kernels.
static constexpr bool	sc_haveNbnxmSimd4xmKernels = 0
	Whether we have support for NBNxM 4xM kernels.

Enumeration Type Documentation

enum KernelLayout

strong

The types of nbNxM SIMD kernel layout.

Enumerator
r4xM	4 'i'-registers each containing data for interaction with M j-atoms
r2xMM	2 'i'-registers each containing duplicated data, { M, M }, for interaction with M j-atoms

enum KernelLayoutClusterRatio

strong

List of supported ratios for j-cluster size versus i-cluster sizes.

Enumerator
JSizeEqualsISize	j-cluster size = i-cluster size
JSizeIsDoubleISize	j-cluster size = 2 * i-cluster size
JSizeIsHalfISize	j-cluster size = i-cluster size / 2