Skip to content

Latest commit

 

History

History
297 lines (219 loc) · 10.8 KB

README.md

File metadata and controls

297 lines (219 loc) · 10.8 KB
Raw

Grid-SIMD

Travis status

Stripped-down version of the Grid: Data parallel C++ mathematical object library only containing the SIMD vectorisation tools.

License: GPL v2.

Description

This library provides data parallel C++ container classes with internal memory layout that is transformed to map efficiently to SIMD architectures. CSHIFT facilities are provided, similar to HPF and cmfortran, and user control is given over the mapping of array indices to both MPI tasks and SIMD processing elements.

  • Identically shaped arrays then be processed with perfect data parallelisation.
  • Such identically shaped arrays are called conformable arrays.

The transformation is based on the observation that Cartesian array processing involves identical processing to be performed on different regions of the Cartesian array.

The library will both geometrically decompose into MPI tasks and across SIMD lanes. Local vector loops are parallelised with OpenMP pragmas.

Data parallel array operations can then be specified with a SINGLE data parallel paradigm, but optimally use MPI, OpenMP and SIMD parallelism under the hood. This is a significant simplification for most programmers.

The layout transformations are parametrised by the SIMD vector length. This adapts according to the architecture. Presently SSE4, ARM NEON (128 bits) AVX, AVX2, QPX (256 bits), IMCI and AVX512 (512 bits) targets are supported.

These are presented as vRealF, vRealD, vComplexF, and vComplexD internal vector data types. The corresponding scalar types are named RealF, RealD, ComplexF and ComplexD.

MPI, OpenMP, and SIMD parallelism are present in the library. Please see this paper for more detail.

Compilers

Intel ICPC v16.0.3 and later

Clang v3.5 and later (need 3.8 and later for OpenMP)

GCC v4.9.x (recommended)

GCC v6.3 and later

Important:

Some versions of GCC appear to have a bug under high optimisation (-O2, -O3).

The safety of these compiler versions cannot be guaranteed at this time. Follow Issue 100 for details and updates.

GCC v5.x

GCC v6.1, v6.2

Possible SIMD types

The following options can be use with the --enable-simd= option to target different SIMD instruction sets:

<code> Description
GEN generic portable vector code
SSE4 SSE 4.2 (128 bit)
AVX AVX (256 bit)
AVXFMA AVX (256 bit) + FMA
AVXFMA4 AVX (256 bit) + FMA4
AVX2 AVX 2 (256 bit)
AVX512 AVX 512 bit
NEONv8 ARM NEON (128 bit)
QPX IBM QPX (256 bit)

Alternatively, some CPU codenames can be directly used:

<code> Description
KNL Intel Xeon Phi codename Knights Landing
SKL Intel Skylake with AVX512 extensions
BGQ Blue Gene/Q

Notes:

  • We currently support AVX512 for the Intel compiler and GCC (KNL and SKL target). Support for clang will appear in future versions of Grid when the AVX512 support in the compiler will be more advanced.
  • For BG/Q only bgclang is supported. We do not presently plan to support more compilers for this platform.
  • BG/Q performances are currently rather poor. This is being investigated for future versions.
  • The vector size for the GEN target can be specified with the configure script option --enable-gen-simd-width.

Build setup for Intel Knights Landing platform

The following configuration is recommended for the Intel Knights Landing platform:

../configure --enable-precision=double\
             --enable-simd=KNL        \
             --enable-comms=mpi-auto  \
             --enable-mkl             \
             CXX=icpc MPICXX=mpiicpc

The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.

If you are working on a Cray machine that does not use the mpiicpc wrapper, please use:

../configure --enable-precision=double\
             --enable-simd=KNL        \
             --enable-comms=mpi       \
             --enable-mkl             \
             CXX=CC CC=cc

If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:

               --with-gmp=<path>        \
               --with-mpfr=<path>       \

where <path> is the UNIX prefix where GMP and MPFR are installed.

Knight's Landing with Intel Omnipath adapters with two adapters per node presently performs better with use of more than one rank per node, using shared memory for interior communication. This is the mpi3 communications implementation. We recommend four ranks per node for best performance, but optimum is local volume dependent.

../configure --enable-precision=double\
             --enable-simd=KNL        \
             --enable-comms=mpi3-auto \
             --enable-mkl             \
             CC=icpc MPICXX=mpiicpc

Build setup for Intel Haswell Xeon platform

The following configuration is recommended for the Intel Haswell platform:

../configure --enable-precision=double\
             --enable-simd=AVX2       \
             --enable-comms=mpi3-auto \
             --enable-mkl             \
             CXX=icpc MPICXX=mpiicpc

The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.

If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:

               --with-gmp=<path>        \
               --with-mpfr=<path>       \

where <path> is the UNIX prefix where GMP and MPFR are installed.

If you are working on a Cray machine that does not use the mpiicpc wrapper, please use:

../configure --enable-precision=double\
             --enable-simd=AVX2       \
             --enable-comms=mpi3      \
             --enable-mkl             \
             CXX=CC CC=cc

Since Dual socket nodes are commonplace, we recommend MPI-3 as the default with the use of one rank per socket. If using the Intel MPI library, threads should be pinned to NUMA domains using

        export I_MPI_PIN=1

This is the default.

Build setup for Intel Skylake Xeon platform

The following configuration is recommended for the Intel Skylake platform:

../configure --enable-precision=double\
             --enable-simd=AVX512     \
             --enable-comms=mpi3      \
             --enable-mkl             \
             CXX=mpiicpc

The MKL flag enables use of BLAS and FFTW from the Intel Math Kernels Library.

If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:

               --with-gmp=<path>        \
               --with-mpfr=<path>       \

where <path> is the UNIX prefix where GMP and MPFR are installed.

If you are working on a Cray machine that does not use the mpiicpc wrapper, please use:

../configure --enable-precision=double\
             --enable-simd=AVX512     \
             --enable-comms=mpi3      \
             --enable-mkl             \
             CXX=CC CC=cc

Since Dual socket nodes are commonplace, we recommend MPI-3 as the default with the use of one rank per socket. If using the Intel MPI library, threads should be pinned to NUMA domains using

        export I_MPI_PIN=1

This is the default.

Expected Skylake Gold 6148 dual socket (single prec, single node 20+20 cores) performance using NUMA MPI mapping):

mpirun -n 2 benchmarks/Benchmark_dwf --grid 16.16.16.16 --mpi 2.1.1.1 --cacheblocking 2.2.2.2 --dslash-asm --shm 1024 --threads 18

TBA

Build setup for AMD EPYC / RYZEN

The AMD EPYC is a multichip module comprising 32 cores spread over four distinct chips each with 8 cores. So, even with a single socket node there is a quad-chip module. Dual socket nodes with 64 cores total are common. Each chip within the module exposes a separate NUMA domain. There are four NUMA domains per socket and we recommend one MPI rank per NUMA domain. MPI-3 is recommended with the use of four ranks per socket, and 8 threads per rank.

The following configuration is recommended for the AMD EPYC platform.

../configure --enable-precision=double\
             --enable-simd=AVX2       \
             --enable-comms=mpi3 \
             CXX=mpicxx

If gmp and mpfr are NOT in standard places (/usr/) these flags may be needed:

               --with-gmp=<path>        \
               --with-mpfr=<path>       \

where <path> is the UNIX prefix where GMP and MPFR are installed.

Using MPICH and g++ v4.9.2, best performance can be obtained using explicit GOMP_CPU_AFFINITY flags for each MPI rank. This can be done by invoking MPI on a wrapper script omp_bind.sh to handle this.

It is recommended to run 8 MPI ranks on a single dual socket AMD EPYC, with 8 threads per rank using MPI3 and shared memory to communicate within this node:

mpirun -np 8 ./omp_bind.sh ./Benchmark_dwf --mpi 2.2.2.1 --dslash-unroll --threads 8 --grid 16.16.16.16 --cacheblocking 4.4.4.4

Where omp_bind.sh does the following:

#!/bin/bash

numanode=` expr $PMI_RANK % 8 `
basecore=`expr $numanode \* 16`
core0=`expr $basecore + 0 `
core1=`expr $basecore + 2 `
core2=`expr $basecore + 4 `
core3=`expr $basecore + 6 `
core4=`expr $basecore + 8 `
core5=`expr $basecore + 10 `
core6=`expr $basecore + 12 `
core7=`expr $basecore + 14 `

export GOMP_CPU_AFFINITY="$core0 $core1 $core2 $core3 $core4 $core5 $core6 $core7"
echo GOMP_CUP_AFFINITY $GOMP_CPU_AFFINITY

$@

Performance:

Expected AMD EPYC 7601 dual socket (single prec, single node 32+32 cores) performance using NUMA MPI mapping):

mpirun -np 8 ./omp_bind.sh ./Benchmark_dwf --threads 8 --mpi 2.2.2.1 --dslash-unroll --grid 16.16.16.16 --cacheblocking 4.4.4.4

TBA

Build setup for BlueGene/Q

To be written...

Build setup for ARM Neon

To be written...

Build setup for laptops, other compilers, non-cluster builds

Many versions of g++ and clang++ work with Grid, and involve merely replacing CXX (and MPICXX), and omit the enable-mkl flag.

Single node builds are enabled with

            --enable-comms=none

FFTW support that is not in the default search path may then enabled with

    --with-fftw=<installpath>

BLAS will not be compiled in by default, and Lanczos will default to Eigen diagonalisation.