From 249165d1b29c3fc38f6f85655a3051dba7ded578 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Tue, 5 May 2015 18:09:28 +0100
Subject: [PATCH] Added streaming stores

---
 lib/simd/Grid_vComplexD.h | 15 +++++++++++++++
 lib/simd/Grid_vComplexF.h | 29 ++++++++++++++++++++++-------
 lib/simd/Grid_vInteger.h  |  4 ++++
 lib/simd/Grid_vRealD.h    | 15 +++++++++++++++
 lib/simd/Grid_vRealF.h    | 15 +++++++++++++++
 5 files changed, 71 insertions(+), 7 deletions(-)

diff --git a/lib/simd/Grid_vComplexD.h b/lib/simd/Grid_vComplexD.h
index 208e2640..f0108d59 100644
--- a/lib/simd/Grid_vComplexD.h
+++ b/lib/simd/Grid_vComplexD.h
@@ -242,6 +242,21 @@ friend inline void vstore(const vComplexD &ret, ComplexD *a){
 	assert(0);
 #endif
         }
+        friend inline void vstream(vComplexD &out,const vComplexD &in){
+#if defined (AVX1)|| defined (AVX2)
+	  _mm256_stream_pd((double *)&out.v,in.v);
+#endif
+#ifdef SSE4
+	  _mm_stream_pd((double *)&out.v,in.v);
+#endif
+#ifdef AVX512
+	  _mm512_stream_pd((double *)&out.v,in.v);
+	  //Note v has a3 a2 a1 a0
+#endif
+#ifdef QPX
+	  assert(0);
+#endif
+	}
       friend inline void vprefetch(const vComplexD &v)
         {
             _mm_prefetch((const char*)&v.v,_MM_HINT_T0);
diff --git a/lib/simd/Grid_vComplexF.h b/lib/simd/Grid_vComplexF.h
index 9c7922e3..5f52fc53 100644
--- a/lib/simd/Grid_vComplexF.h
+++ b/lib/simd/Grid_vComplexF.h
@@ -176,21 +176,36 @@ namespace Grid {
             vsplat(ret,a,b);
         }
 
-friend inline void vstore(const vComplexF &ret, ComplexF *a){
+	friend inline void vstore(const vComplexF &ret, ComplexF *a){
 #if defined (AVX1)|| defined (AVX2)
-        _mm256_store_ps((float *)a,ret.v);
+	  _mm256_store_ps((float *)a,ret.v);
 #endif
 #ifdef SSE4
-        _mm_store_ps((float *)a,ret.v);
+	  _mm_store_ps((float *)a,ret.v);
 #endif
 #ifdef AVX512
-	_mm512_store_ps((float *)a,ret.v);
-//Note v has a3 a2 a1 a0
+	  _mm512_store_ps((float *)a,ret.v);
+	  //Note v has a3 a2 a1 a0
 #endif
 #ifdef QPX
-	assert(0);
+	  assert(0);
 #endif
-}
+	}
+        friend inline void vstream(vComplexF &out,const vComplexF &in){
+#if defined (AVX1)|| defined (AVX2)
+	  _mm256_stream_ps((float *)&out.v,in.v);
+#endif
+#ifdef SSE4
+	  _mm_stream_ps((float *)&out.v,in.v);
+#endif
+#ifdef AVX512
+	  _mm512_stream_ps((float *)&out.v,in.v);
+	  //Note v has a3 a2 a1 a0
+#endif
+#ifdef QPX
+	  assert(0);
+#endif
+	}
       friend inline void vprefetch(const vComplexF &v)
         {
             _mm_prefetch((const char*)&v.v,_MM_HINT_T0);
diff --git a/lib/simd/Grid_vInteger.h b/lib/simd/Grid_vInteger.h
index 6035aea1..5a429a9c 100644
--- a/lib/simd/Grid_vInteger.h
+++ b/lib/simd/Grid_vInteger.h
@@ -186,6 +186,10 @@ namespace Grid {
 #endif
         }
 
+	friend inline void vstream(vInteger & out,const vInteger &in){
+	  out=in;
+        }
+
         friend inline void vprefetch(const vInteger &v)
         {
             _mm_prefetch((const char*)&v.v,_MM_HINT_T0);
diff --git a/lib/simd/Grid_vRealD.h b/lib/simd/Grid_vRealD.h
index 0bdaa5e4..ee51fa03 100644
--- a/lib/simd/Grid_vRealD.h
+++ b/lib/simd/Grid_vRealD.h
@@ -174,6 +174,21 @@ namespace Grid {
 #endif
 #ifdef QPX
 	    assert(0);
+#endif
+	}
+        friend inline void vstream(vRealD &out,const vRealD &in){
+#if defined (AVX1)|| defined (AVX2)
+	  _mm256_stream_pd((double *)&out.v,in.v);
+#endif
+#ifdef SSE4
+	  _mm_stream_pd((double *)&out.v,in.v);
+#endif
+#ifdef AVX512
+	  _mm512_stream_pd((double *)&out.v,in.v);
+	  //Note v has a3 a2 a1 a0
+#endif
+#ifdef QPX
+	  assert(0);
 #endif
 	}
         friend inline void vprefetch(const vRealD &v)
diff --git a/lib/simd/Grid_vRealF.h b/lib/simd/Grid_vRealF.h
index e8d89cec..48e5c134 100644
--- a/lib/simd/Grid_vRealF.h
+++ b/lib/simd/Grid_vRealF.h
@@ -208,6 +208,21 @@ friend inline void vstore(const vRealF &ret, float *a){
 	assert(0);
 #endif
         }
+        friend inline void vstream(vRealF &out,const vRealF &in){
+#if defined (AVX1)|| defined (AVX2)
+	  _mm256_stream_ps((float *)&out.v,in.v);
+#endif
+#ifdef SSE4
+	  _mm_stream_ps((float *)&out.v,in.v);
+#endif
+#ifdef AVX512
+	  _mm512_stream_ps((float *)&out.v,in.v);
+	  //Note v has a3 a2 a1 a0
+#endif
+#ifdef QPX
+	  assert(0);
+#endif
+	}
 
 
         friend inline void vprefetch(const vRealF &v)