From 661fc4d3d1f2555bf58f5c5dd2ef876435a0cc37 Mon Sep 17 00:00:00 2001
From: paboyle <paboyle@ph.ed.ac.uk>
Date: Mon, 20 Feb 2017 17:48:36 -0500
Subject: [PATCH] Debug AVX512 exchange code paths

---
 lib/simd/Grid_avx512.h       |  6 ++++++
 lib/simd/Grid_vector_types.h | 17 +++++++++++++----
 2 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/lib/simd/Grid_avx512.h b/lib/simd/Grid_avx512.h
index 22d45aeb..f39c4033 100644
--- a/lib/simd/Grid_avx512.h
+++ b/lib/simd/Grid_avx512.h
@@ -355,6 +355,8 @@ namespace Optimization {
     static inline void Exchange1(__m512 &out1,__m512 &out2,__m512 in1,__m512 in2){
       out1= _mm512_shuffle_f32x4(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0));
       out2= _mm512_shuffle_f32x4(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1));
+      out1= _mm512_shuffle_f32x4(out1,out1,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
+      out2= _mm512_shuffle_f32x4(out2,out2,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
     };
     static inline void Exchange2(__m512 &out1,__m512 &out2,__m512 in1,__m512 in2){
       out1= _mm512_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(1,0,1,0));
@@ -363,6 +365,8 @@ namespace Optimization {
     static inline void Exchange3(__m512 &out1,__m512 &out2,__m512 in1,__m512 in2){
       out1= _mm512_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0));
       out2= _mm512_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1));
+      out1= _mm512_shuffle_ps(out1,out1,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
+      out2= _mm512_shuffle_ps(out2,out2,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
     };
  
     static inline void Exchange0(__m512d &out1,__m512d &out2,__m512d in1,__m512d in2){
@@ -372,6 +376,8 @@ namespace Optimization {
     static inline void Exchange1(__m512d &out1,__m512d &out2,__m512d in1,__m512d in2){
       out1= _mm512_shuffle_f64x2(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0));
       out2= _mm512_shuffle_f64x2(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1));
+      out1= _mm512_shuffle_f64x2(out1,out1,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
+      out2= _mm512_shuffle_f64x2(out2,out2,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
     };
     static inline void Exchange2(__m512d &out1,__m512d &out2,__m512d in1,__m512d in2){
       out1 = _mm512_shuffle_pd(in1,in2,0x00);
diff --git a/lib/simd/Grid_vector_types.h b/lib/simd/Grid_vector_types.h
index cd499d88..dcdacbe0 100644
--- a/lib/simd/Grid_vector_types.h
+++ b/lib/simd/Grid_vector_types.h
@@ -356,10 +356,19 @@ class Grid_simd {
   ///////////////////////
   friend inline void exchange(Grid_simd &out1,Grid_simd &out2,Grid_simd in1,Grid_simd in2,int n)
   {
-    if     (n==3) Optimization::Exchange::Exchange3(out1.v,out2.v,in1.v,in2.v);
-    else if(n==2) Optimization::Exchange::Exchange2(out1.v,out2.v,in1.v,in2.v);
-    else if(n==1) Optimization::Exchange::Exchange1(out1.v,out2.v,in1.v,in2.v);
-    else if(n==0) Optimization::Exchange::Exchange0(out1.v,out2.v,in1.v,in2.v);
+    if       (n==3) {
+      Optimization::Exchange::Exchange3(out1.v,out2.v,in1.v,in2.v);
+      //      std::cout << " Exchange3 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
+    } else if(n==2) {
+      Optimization::Exchange::Exchange2(out1.v,out2.v,in1.v,in2.v);
+      //      std::cout << " Exchange2 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
+    } else if(n==1) {
+      Optimization::Exchange::Exchange1(out1.v,out2.v,in1.v,in2.v);
+      //      std::cout << " Exchange1 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
+    } else if(n==0) { 
+      Optimization::Exchange::Exchange0(out1.v,out2.v,in1.v,in2.v);
+      //      std::cout << " Exchange0 "<< out1<<" "<< out2<<" <- " << in1 << " "<<in2<<std::endl;
+    }
   }
 
   ////////////////////////////////////////////////////////////////////