From fbed02690d8a3ab1889f769afe19954fb99698e1 Mon Sep 17 00:00:00 2001
From: nmeyer-ur <nils.meyer@ur.de>
Date: Fri, 8 May 2020 12:05:31 +0200
Subject: [PATCH] some changes in breaking out A64FX: use -DA64FXFIXEDSIZE for
 fixed size, but also define GEN

---
 Grid/simd/Grid_a64fx-fixedsize.h |  4 +++-
 Grid/simd/Grid_vector_types.h    | 20 ++++++++++++--------
 2 files changed, 15 insertions(+), 9 deletions(-)

diff --git a/Grid/simd/Grid_a64fx-fixedsize.h b/Grid/simd/Grid_a64fx-fixedsize.h
index 34ba955c..901c85b8 100644
--- a/Grid/simd/Grid_a64fx-fixedsize.h
+++ b/Grid/simd/Grid_a64fx-fixedsize.h
@@ -36,6 +36,7 @@
 
 //static_assert(GEN_SIMD_WIDTH % 64u == 0, "A64FX SIMD vector size is 64 bytes");
 
+/*
 #ifdef __ARM_FEATURE_SVE
   #pragma message("Yes, we have SVE feature")
   #include <arm_sve.h>
@@ -43,6 +44,7 @@
   #pragma error "Missing SVE feature"
 #endif /* __ARM_FEATURE_SVE */
 
+
 // gcc 10 features
 #if __ARM_FEATURE_SVE_BITS==512
 #pragma message("Fixed-size SVE ACLE")
@@ -58,7 +60,7 @@ typedef svuint64_t  lutd __attribute__((arm_sve_vector_bits(512))); // LUTs for
 #endif /* __ARM_FEATURE_SVE_BITS */
 
 // safety definition, not sure if it's necessary
-#define GEN_SIMD_WIDTH 64u
+//#define GEN_SIMD_WIDTH 64u
 
 // low-level API
 NAMESPACE_BEGIN(Grid);
diff --git a/Grid/simd/Grid_vector_types.h b/Grid/simd/Grid_vector_types.h
index 55d521b3..f8223441 100644
--- a/Grid/simd/Grid_vector_types.h
+++ b/Grid/simd/Grid_vector_types.h
@@ -119,25 +119,29 @@ accelerator_inline Grid_half sfw_float_to_half(float ff) {
 #ifdef A64FXGCC
   #include "Grid_a64fx-fixedsize.h"
 #else
-  #ifdef GEN
-    #if defined(A64FX) // breakout A64FX SVE ACLE here
+
+#ifdef GEN
+  #if defined(A64FX) || defined(A64FXGCC) // breakout A64FX SVE ACLE here
+    #include <arm_sve.h>
+    if defined(A64FX) // VLA
       #pragma message("building for A64FX / SVE ACLE VLA")
       #if defined(ARMCLANGCOMPAT)
         #pragma message("applying armclang fix")
       //#else
-      //  #pragma message("not applying armclang fix")
+        //#pragma message("not applying armclang fix")
       #endif
-      #include <arm_sve.h>
       #include "Grid_a64fx-2.h"
-    #else
-      #pragma message("building for A64FX / GEN")
-      #include "Grid_generic.h"
+    #endif
+    #if defined(A64FXFIXEDSIZE) // fixed size data types
+      #include "Grid_a64fx-fixedsize.h"
     #endif
   #else
-    #pragma error("Undefined architecture")
+    #pragma message("building for GEN") // generic
+    #include "Grid_generic.h"
   #endif
 #endif
 
+
 #ifdef SSE4
 #include "Grid_sse4.h"
 #endif