From 046b1cbbc0b771f33979b124ba0a878d8f1e5cf7 Mon Sep 17 00:00:00 2001
From: nmeyer-ur <nils.meyer@ur.de>
Date: Thu, 21 May 2020 19:39:07 +0200
Subject: [PATCH] enable fcmla in tensor arithmetics; fixed-size works, VLA
 does not compile

---
 Grid/simd/Grid_a64fx-2.h         | 18 ++++++++++++++++--
 Grid/simd/Grid_a64fx-fixedsize.h | 16 ++++++++--------
 2 files changed, 24 insertions(+), 10 deletions(-)

diff --git a/Grid/simd/Grid_a64fx-2.h b/Grid/simd/Grid_a64fx-2.h
index 0333299f..65254e90 100644
--- a/Grid/simd/Grid_a64fx-2.h
+++ b/Grid/simd/Grid_a64fx-2.h
@@ -324,6 +324,18 @@ struct Sub{
 };
 
 struct Mult{
+  template <typename T>
+  inline vec<T> operator()(vec<T> a, vec<T> b, vec<T> c){
+    vec<T> out;
+    svbool_t pg1 = acle<T>::pg1();
+    typename acle<T>::vt a_v = svld1(pg1, a.v);
+    typename acle<T>::vt b_v = svld1(pg1, b.v);
+    typename acle<T>::vt c_v = svld1(pg1, c.v);
+    typename acle<T>::vt r_v = svmad_x(pg1, b_v, c_v, a_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
+  }
   template <typename T>
   inline vec<T> operator()(vec<T> a, vec<T> b){
     vec<T> out;
@@ -396,7 +408,7 @@ struct MultComplex{
 struct MultAddComplex{
   // Complex a*b+c
   template <typename T>
-  inline void mac(const vec<T> &a, const vec<T> b, const vec<T> c){
+  inline vec<T> operator()(vec<T> a, vec<T> b, vec<T> c){
     vec<T> out;
     svbool_t pg1 = acle<T>::pg1();
     typename acle<T>::vt a_v = svld1(pg1, a.v);
@@ -407,7 +419,9 @@ struct MultAddComplex{
     typename acle<T>::vt r_v = svcmla_x(pg1, c_v, a_v, b_v, 0);
     r_v = svcmla_x(pg1, r_v, a_v, b_v, 90);
 
-    svst1(pg1, a.v, r_v);
+    svst1(pg1, out.v, r_v);
+
+    return out;
   }
 };
 
diff --git a/Grid/simd/Grid_a64fx-fixedsize.h b/Grid/simd/Grid_a64fx-fixedsize.h
index a07cbff3..b3b93884 100644
--- a/Grid/simd/Grid_a64fx-fixedsize.h
+++ b/Grid/simd/Grid_a64fx-fixedsize.h
@@ -295,14 +295,14 @@ struct Sub{
 
 struct Mult{
   // Real float fma
-  inline void mac(vecf &a, vecf b, vecf c){
+  inline void operator()(vecf a, vecf b, vecf c){
     pred pg1 = acle<float>::pg1();
-    a = svmad_x(pg1, b, c, a);
+    return svmad_x(pg1, b, c, a);
   }
   // Real double fma
-  inline void mac(vecd &a, vecd b, vecd c){
+  inline void operator()(vecd a, vecd b, vecd c){
     pred pg1 = acle<double>::pg1();
-    a = svmad_x(pg1, b, c, a);
+    return svmad_x(pg1, b, c, a);
   }
   // Real float
   inline vecf operator()(vecf a, vecf b){
@@ -376,18 +376,18 @@ struct MultComplex{
 struct MultAddComplex{
   // Complex a*b+c
   // Complex float
-  inline void mac(vecf &a, vecf b, vecf c){
+  inline vecf operator()(vecf a, vecf b, vecf c){
     pred pg1 = acle<float>::pg1();
     // using FCMLA
     vecf r_v = svcmla_x(pg1, c, a, b, 0);
-    a = svcmla_x(pg1, r_v, a, b, 90);
+    return svcmla_x(pg1, r_v, a, b, 90);
   }
   // Complex double
-  inline void mac(vecd &a, vecd b, vecd c){
+  inline vecd operator()(vecd a, vecd b, vecd c){
     pred pg1 = acle<double>::pg1();
     // using FCMLA
     vecd r_v = svcmla_x(pg1, c, a, b, 0);
-    a = svcmla_x(pg1, r_v, a, b, 90);
+    return svcmla_x(pg1, r_v, a, b, 90);
   }
 };