From bbd145382b0bb866f3c5859d603eb6946bfe369c Mon Sep 17 00:00:00 2001
From: nmeyer-ur <nils.meyer@ur.de>
Date: Wed, 8 Jul 2020 12:43:51 +0200
Subject: [PATCH] enable --enable-simd=A64FX in configure

---
 .../implementation/WilsonKernelsAsmA64FX.h    |  7 ++--
 Grid/simd/Grid_vector_types.h                 | 34 +++++++++++++++++--
 configure.ac                                  |  9 +++++
 3 files changed, 45 insertions(+), 5 deletions(-)

diff --git a/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h b/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h
index 34da3110..2e587dfa 100644
--- a/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h
+++ b/Grid/qcd/action/fermion/implementation/WilsonKernelsAsmA64FX.h
@@ -29,7 +29,8 @@ Author: Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 /*  END LEGAL */
 #pragma once
 
-#if defined(A64FXASM)
+//#if defined(A64FXASM)
+#if defined(A64FX)
 
 // safety include
 #include <arm_sve.h>
@@ -39,13 +40,13 @@ Author: Nils Meyer  <nils.meyer@ur.de>  Regensburg University
 
 // enable A64FX body
 #define WILSONKERNELSASMBODYA64FX
-#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h")
+//#pragma message("A64FX Dslash: WilsonKernelsAsmBodyA64FX.h")
 
     ///////////////////////////////////////////////////////////
     // If we are A64FX specialise the single precision routine
     ///////////////////////////////////////////////////////////
 #if defined(DSLASHINTRIN)
-#pragma message ("A64FX Dslash: intrin")
+//#pragma message ("A64FX Dslash: intrin")
 #include <simd/Fujitsu_A64FX_intrin_single.h>
 #else
 #pragma message ("A64FX Dslash: asm")
diff --git a/Grid/simd/Grid_vector_types.h b/Grid/simd/Grid_vector_types.h
index 727ed668..33ebe8c1 100644
--- a/Grid/simd/Grid_vector_types.h
+++ b/Grid/simd/Grid_vector_types.h
@@ -120,7 +120,7 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
   #if defined(A64FX) || defined(A64FXFIXEDSIZE) // breakout A64FX SVE ACLE here
     #include <arm_sve.h>
     #if defined(A64FX) // VLA
-      #pragma message("building for A64FX / SVE ACLE VLA")
+      #pragma message("building A64FX / SVE ACLE VLA")
       #if defined(ARMCLANGCOMPAT)
         #pragma message("applying data types patch")
       #endif
@@ -131,11 +131,41 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
       #include "Grid_a64fx-fixedsize.h"
     #endif
   #else
-    #pragma message("building for GEN") // generic
+    #pragma message("building GEN") // generic
     #include "Grid_generic.h"
   #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
 #include "Grid_sse4.h"
diff --git a/configure.ac b/configure.ac
index 878d56e6..f0bea6a4 100644
--- a/configure.ac
+++ b/configure.ac
@@ -392,6 +392,15 @@ case ${ax_cv_cxx_compiler_vendor} in
                            [generic SIMD vector width (in bytes)])
         SIMD_GEN_WIDTH_MSG=" (width= $ac_gen_simd_width)"
         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)
         AC_DEFINE([NEONV8],[1],[ARMv8 NEON])
         SIMD_FLAGS='-march=armv8-a';;