portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-09 21:50:45 +01:00
Code Releases Activity

enable --enable-simd=A64FX in configure

Browse Source
This commit is contained in:
nmeyer-ur 2020-07-08 12:43:51 +02:00
parent 1b08cb7300
commit bbd145382b
3 changed files with 45 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h
View File

@ -29,7 +29,8 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
/* END LEGAL */ /* END LEGAL */
#pragma once #pragma once
#if defined(A64FXASM) //#if defined(A64FXASM)
#if defined(A64FX)
// safety include // safety include
#include <arm_sve.h> #include <arm_sve.h>
@ -39,13 +40,13 @@ Author: Nils Meyer <nils.meyer@ur.de> Regensburg University
// enable A64FX body // enable A64FX body
#define WILSONKERNELSASMBODYA64FX #define WILSONKERNELSASMBODYA64FX
#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h") //#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h")
/////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////
// If we are A64FX specialise the single precision routine // If we are A64FX specialise the single precision routine
/////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////
#if defined(DSLASHINTRIN) #if defined(DSLASHINTRIN)
#pragma message ("A64FX Dslash: intrin") //#pragma message ("A64FX Dslash: intrin")
#include <simd/Fujitsu_A64FX_intrin_single.h> #include <simd/Fujitsu_A64FX_intrin_single.h>
#else #else
#pragma message ("A64FX Dslash: asm") #pragma message ("A64FX Dslash: asm")

34
Grid/simd/Grid_vector_types.h
View File

@ -120,7 +120,7 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
#if defined(A64FX) || defined(A64FXFIXEDSIZE) // breakout A64FX SVE ACLE here #if defined(A64FX) || defined(A64FXFIXEDSIZE) // breakout A64FX SVE ACLE here
#include <arm_sve.h> #include <arm_sve.h>
#if defined(A64FX) // VLA #if defined(A64FX) // VLA
#pragma message("building for A64FX / SVE ACLE VLA") #pragma message("building A64FX / SVE ACLE VLA")
#if defined(ARMCLANGCOMPAT) #if defined(ARMCLANGCOMPAT)
#pragma message("applying data types patch") #pragma message("applying data types patch")
#endif #endif
@ -131,11 +131,41 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
#include "Grid_a64fx-fixedsize.h" #include "Grid_a64fx-fixedsize.h"
#endif #endif
#else #else
#pragma message("building for GEN") // generic #pragma message("building GEN") // generic
#include "Grid_generic.h" #include "Grid_generic.h"
#endif #endif
#endif #endif
#ifdef A64FX
#include <arm_sve.h>
#ifdef __ARM_FEATURE_SVE_BITS
//#pragma message("building A64FX SVE VLS")
#include "Grid_a64fx-fixedsize.h"
#else
#pragma message("building A64FX SVE VLA")
#if defined(ARMCLANGCOMPAT)
#pragma message("applying data types patch")
#endif
#include "Grid_a64fx-2.h"
#endif
#endif
/*
#ifdef A64FXVLA
#pragma message("building A64FX VLA")
#if defined(ARMCLANGCOMPAT)
#pragma message("applying data types patch")
#endif
#include <arm_sve.h>
#include "Grid_a64fx-2.h"
#endif
#ifdef A64FXVLS
#pragma message("building A64FX VLS")
#include <arm_sve.h>
#include "Grid_a64fx-fixedsize.h"
#endif
*/
#ifdef SSE4 #ifdef SSE4
#include "Grid_sse4.h" #include "Grid_sse4.h"

9
configure.ac
View File

@ -392,6 +392,15 @@ case ${ax_cv_cxx_compiler_vendor} in
[generic SIMD vector width (in bytes)]) [generic SIMD vector width (in bytes)])
SIMD_GEN_WIDTH_MSG=" (width= $ac_gen_simd_width)" SIMD_GEN_WIDTH_MSG=" (width= $ac_gen_simd_width)"
SIMD_FLAGS='';; SIMD_FLAGS='';;
A64FX)
case ${ax_cv_cxx_compiler_vendor} in
gnu)
AC_DEFINE([A64FX],[1],[A64FX / 512-bit SVE VLS])
SIMD_FLAGS='-march=armv8.2-a+sve -msve-vector-bits=512 -fno-gcse -DDSLASHINTRIN';;
clang)
AC_DEFINE([A64FX],[1],[A64FX / 512-bit SVE VLA])
SIMD_FLAGS='-mcpu=a64fx -DARMCLANGCOMPAT -DDSLASHINTRIN';;
esac;;
NEONv8) NEONv8)
AC_DEFINE([NEONV8],[1],[ARMv8 NEON]) AC_DEFINE([NEONV8],[1],[ARMv8 NEON])
SIMD_FLAGS='-march=armv8-a';; SIMD_FLAGS='-march=armv8-a';;