From b2933a0557d211fa3b3d86d9d9ee4d38ebb854b1 Mon Sep 17 00:00:00 2001
From: paboyle <paboyle@ph.ed.ac.uk>
Date: Sat, 25 Jun 2016 12:55:25 -0700
Subject: [PATCH] COntrol the prefetch strategy

---
 lib/simd/Intel512common.h | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/lib/simd/Intel512common.h b/lib/simd/Intel512common.h
index 6878bcfb..a05f978c 100644
--- a/lib/simd/Intel512common.h
+++ b/lib/simd/Intel512common.h
@@ -28,6 +28,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #ifndef GRID_ASM_INTEL_COMMON_512_H
 #define GRID_ASM_INTEL_COMMON_512_H
 
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// Peformance options
+////////////////////////////////////////////////////////////////////////////////////////////////////
+#define AVX512_PF_L1
+#undef  AVX512_PF_L2_LINEAR
+#undef  AVX512_PF_L2_TABLE
+#undef  AVX512_PF_L2_WRITE
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // Opcodes common 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -88,10 +96,30 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
 #define VMOVf(A,DEST)   "vmovaps  " #A ", " #DEST  ";\n"
 #define VMOVd(A,DEST)   "vmovapd  " #A ", " #DEST  ";\n"
 
+#ifdef AVX512_PF_L1
 #define VPREFETCHG(O,A) "prefetcht0 "#O"*64("#A");\n" 
+#else
+#define VPREFETCHG(O,A) 
+#endif
+
+#ifdef AVX512_PF_L2_LINEAR
 #define VPREFETCH2(O,A) "prefetcht1 "#O"*64("#A");\n" 
+#else
+#define VPREFETCH2(O,A) 
+#endif
+
+#ifdef AVX512_PF_L2_TABLE
 #define VPREFETCHP(O,A) "prefetcht1 "#O"*64("#A");\n" 
+#else
+#define VPREFETCHP(O,A) 
+#endif
+
+#ifdef AVX512_PF_L2_WRITE
 #define VPREFETCHW(O,A) "prefetchwt1 "#O"*64("#A");\n" 
+#else
+#define VPREFETCHW(O,A) 
+#endif
+
 #define VPREFETCHNTA(O,A) 
 #define VPREFETCH(O,A)