From 8ac87c1cb63fd8518c7a99d6b06fb47524f2153b Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Mon, 16 Oct 2017 23:27:51 +0200
Subject: [PATCH 1/4] Implement DSDOT with unchanged sdot microkernels

---
 kernel/x86_64/sdot.c | 23 ++++++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/kernel/x86_64/sdot.c b/kernel/x86_64/sdot.c
index 389252f8b5..f786d18953 100644
--- a/kernel/x86_64/sdot.c
+++ b/kernel/x86_64/sdot.c
@@ -68,7 +68,11 @@ static void sdot_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
 
 #endif
 
+#if defined (DSDOT)
+double CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+#else
 FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+#endif
 {
 	BLASLONG i=0;
 	BLASLONG ix=0,iy=0;
@@ -91,12 +95,19 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 		i = n1;
 		while(i < n)
 		{
-
+#if defined(DSDOT)
+			dot += (double)y[i] * (double)x[i] ;
+#else
 			dot += y[i] * x[i] ;
+#endif
 			i++ ;
 
 		}
+#if defined(DSDOT)
+                dot+=(double)mydot;
+#else                
 		dot+=mydot;
+#endif
 		return(dot);
 
 
@@ -106,8 +117,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 
 	while(i < n1)
 	{
-
+#if defined (DSDOT)
+		dot += (double)y[iy] * (double)x[ix] + (double)y[iy+inc_y] * (double)x[ix+inc_x];
+#else
 		dot += y[iy] * x[ix] + y[iy+inc_y] * x[ix+inc_x];
+#endif
 		ix  += inc_x*2 ;
 		iy  += inc_y*2 ;
 		i+=2 ;
@@ -116,8 +130,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 
 	while(i < n)
 	{
-
+#if defined (DSDOT)
+		dot += (double)y[iy] * (double)x[ix] ;
+#else
 		dot += y[iy] * x[ix] ;
+#endif
 		ix  += inc_x ;
 		iy  += inc_y ;
 		i++ ;

From 28c3fa8950045d658e2c9b604a061927ffb9dc61 Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Mon, 16 Oct 2017 23:29:03 +0200
Subject: [PATCH 2/4] Add dsdot

---
 kernel/x86_64/KERNEL.HASWELL | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index f2e1374d32..848de38dfc 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -24,6 +24,8 @@ DDOTKERNEL = ddot.c
 CDOTKERNEL = cdot.c
 ZDOTKERNEL = zdot.c
 
+DSDOTKERNEL = sdot.c
+
 SAXPYKERNEL = saxpy.c
 DAXPYKERNEL = daxpy.c
 CAXPYKERNEL = caxpy.c

From 5e3e91d0fc5562782ddac9c01d6765cb24f171a6 Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Sun, 22 Oct 2017 18:18:51 +0200
Subject: [PATCH 3/4] Split the microkernel workload into chunks of 32 floats
 for dsdot mode to limit loss of precision

---
 kernel/x86_64/sdot.c | 28 ++++++++++++++++++++++------
 1 file changed, 22 insertions(+), 6 deletions(-)

diff --git a/kernel/x86_64/sdot.c b/kernel/x86_64/sdot.c
index f786d18953..b6f3c21afe 100644
--- a/kernel/x86_64/sdot.c
+++ b/kernel/x86_64/sdot.c
@@ -78,7 +78,12 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 	BLASLONG ix=0,iy=0;
 	double dot = 0.0 ;
 
+#if defined (DSDOT)
+        double mydot = 0.0;
+        FLOAT asmdot = 0.0;
+#else
 	FLOAT mydot=0.0;
+#endif
 	BLASLONG n1;
 
 	if ( n <= 0 )  return(dot);
@@ -89,9 +94,23 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 	        n1 = n & (BLASLONG)(-32);
 
 		if ( n1 )
+#if defined(DSDOT)
+			{
+			FLOAT *x1=x;
+			FLOAT *y1=y;
+			BLASLONG n2 = 32;
+			while (i<n1) {
+				sdot_kernel_16(n2, x1, y1 , &asmdot );
+				mydot += (double)asmdot;
+				asmdot=0.;
+				x1+=32;
+				y1+=32;
+				i+=32;
+			}
+		}
+#else		
 			sdot_kernel_16(n1, x, y , &mydot );
-
-
+#endif
 		i = n1;
 		while(i < n)
 		{
@@ -103,11 +122,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 			i++ ;
 
 		}
-#if defined(DSDOT)
-                dot+=(double)mydot;
-#else                
+
 		dot+=mydot;
-#endif
 		return(dot);
 
 

From a07807caac631234250a7c1b487812b62265df0e Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Wed, 25 Oct 2017 16:45:41 +0200
Subject: [PATCH 4/4] Eliminate loop code when called as/from dsdot

---
 kernel/x86_64/sdot_microk_haswell-2.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/kernel/x86_64/sdot_microk_haswell-2.c b/kernel/x86_64/sdot_microk_haswell-2.c
index 4051f9c1ba..3248c408cf 100644
--- a/kernel/x86_64/sdot_microk_haswell-2.c
+++ b/kernel/x86_64/sdot_microk_haswell-2.c
@@ -53,9 +53,11 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vfmadd231ps    64(%3,%0,4), %%ymm14, %%ymm6 \n\t"  // 2 * y
 	"vfmadd231ps    96(%3,%0,4), %%ymm15, %%ymm7 \n\t"  // 2 * y
 
+#ifndef DSDOT
 	"addq		$32 , %0	  	     \n\t"
 	"subq	        $32 , %1		     \n\t"		
 	"jnz		1b		             \n\t"
+#endif
 
 	"vextractf128	$1 , %%ymm4 , %%xmm12	     \n\t"
 	"vextractf128	$1 , %%ymm5 , %%xmm13	     \n\t"