From 75f893e145fba9dfdf4ce1321e5b1bf9f78fda97 Mon Sep 17 00:00:00 2001
From: tkelman <tony@kelman.net>
Date: Sat, 21 Jun 2014 22:19:32 -0700
Subject: [PATCH] carry upstream patch for using specific kernels via
 OPENBLAS_CORETYPE

This is cherry-picked from https://github.com/xianyi/OpenBLAS/pull/386
and provides at least a workaround for Haswell and AMD problems seen in #6504, #7031
---
 deps/Makefile               |   1 +
 deps/openblas-dynamic.patch | 163 ++++++++++++++++++++++++++++++++++++
 2 files changed, 164 insertions(+)
 create mode 100644 deps/openblas-dynamic.patch

diff --git a/deps/Makefile b/deps/Makefile
index 76515f945dbef..05e6b1aef2d46 100644
--- a/deps/Makefile
+++ b/deps/Makefile
@@ -804,6 +804,7 @@ endif
 	perl -i -ple 's/^\s*(EXTRALIB\s*\+=\s*-lSystemStubs)\s*$$/# $$1/g' openblas-$(OPENBLAS_VER)/Makefile.system
 ifneq ($(OPENBLAS_VER),develop)
 	patch openblas-$(OPENBLAS_VER)/exports/Makefile < openblas-exports-makefile.patch
+	cd openblas-$(OPENBLAS_VER) && patch -p1 < ../openblas-dynamic.patch
 endif
 	cd openblas-$(OPENBLAS_VER) && patch -p1 < ../openblas-dllinit-safe.patch
 	echo 1 > $@
diff --git a/deps/openblas-dynamic.patch b/deps/openblas-dynamic.patch
new file mode 100644
index 0000000000000..1a7f2057a2f86
--- /dev/null
+++ b/deps/openblas-dynamic.patch
@@ -0,0 +1,163 @@
+diff --git a/driver/others/dynamic.c b/driver/others/dynamic.c
+index 905efb1..ec421d6 100644
+--- a/driver/others/dynamic.c
++++ b/driver/others/dynamic.c
+@@ -116,18 +116,24 @@ extern void openblas_warning(int verbose, const char * msg);
+ 
+ static int get_vendor(void){
+   int eax, ebx, ecx, edx;
+-  char vendor[13];
++
++  union
++  {
++        char vchar[16];
++        int  vint[4];
++  } vendor;
+ 
+   cpuid(0, &eax, &ebx, &ecx, &edx);
+-  
+-  *(int *)(&vendor[0]) = ebx;
+-  *(int *)(&vendor[4]) = edx;
+-  *(int *)(&vendor[8]) = ecx;
+-  vendor[12] = (char)0;
+ 
+-  if (!strcmp(vendor, "GenuineIntel")) return VENDOR_INTEL;
+-  if (!strcmp(vendor, "AuthenticAMD")) return VENDOR_AMD;
+-  if (!strcmp(vendor, "CentaurHauls")) return VENDOR_CENTAUR;
++  *(&vendor.vint[0]) = ebx;
++  *(&vendor.vint[1]) = edx;
++  *(&vendor.vint[2]) = ecx;
++
++  vendor.vchar[12] = '\0';
++
++  if (!strcmp(vendor.vchar, "GenuineIntel")) return VENDOR_INTEL;
++  if (!strcmp(vendor.vchar, "AuthenticAMD")) return VENDOR_AMD;
++  if (!strcmp(vendor.vchar, "CentaurHauls")) return VENDOR_CENTAUR;
+ 
+   if ((eax == 0) || ((eax & 0x500) != 0)) return VENDOR_INTEL;
+ 
+@@ -232,7 +238,7 @@ static gotoblas_t *get_coretype(void){
+     if (family <= 0xe) {
+         // Verify that CPU has 3dnow and 3dnowext before claiming it is Athlon
+         cpuid(0x80000000, &eax, &ebx, &ecx, &edx);
+-        if (eax & 0xffff >= 0x01) {
++        if ( (eax & 0xffff)  >= 0x01) {
+             cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
+             if ((edx & (1 << 30)) == 0 || (edx & (1 << 31)) == 0)
+               return NULL;
+@@ -333,11 +339,79 @@ char *gotoblas_corename(void) {
+   return corename[0];
+ }
+ 
++
++static gotoblas_t *force_coretype(char *coretype){
++
++	int i ;
++	int found = -1;
++	char message[128];
++	char mname[20];
++
++	for ( i=1 ; i <= 20; i++)
++	{
++		if (!strncasecmp(coretype,corename[i],20))
++		{
++			found = i;
++			break;
++		}
++	}
++	if (found < 0)
++	{
++		strncpy(mname,coretype,20);
++    		sprintf(message, "Core not found: %s\n",mname);
++    		openblas_warning(1, message); 
++		return(NULL);
++	}
++
++	switch (found)
++	{
++
++		case 20: return (&gotoblas_HASWELL);
++		case 19: return (&gotoblas_PILEDRIVER);
++		case 18: return (&gotoblas_BULLDOZER);
++		case 17: return (&gotoblas_BOBCAT);
++		case 16: return (&gotoblas_SANDYBRIDGE);
++		case 15: return (&gotoblas_NANO);
++		case 14: return (&gotoblas_BARCELONA);
++		case 13: return (&gotoblas_OPTERON);
++		case 12: return (&gotoblas_OPTERON_SSE3);
++		case 11: return (&gotoblas_ATHLON);
++		case 10: return (&gotoblas_NEHALEM);
++		case  9: return (&gotoblas_DUNNINGTON);
++		case  8: return (&gotoblas_PENRYN);
++		case  7: return (&gotoblas_CORE2);
++		case  6: return (&gotoblas_ATOM);
++		case  5: return (&gotoblas_BANIAS);
++		case  4: return (&gotoblas_PRESCOTT);
++		case  3: return (&gotoblas_NORTHWOOD);
++		case  2: return (&gotoblas_COPPERMINE);
++		case  1: return (&gotoblas_KATMAI);
++	}
++	return(NULL);
++
++}
++			
++	
++		
++
+ void gotoblas_dynamic_init(void) {
+   
++  char coremsg[128];
++  char coren[22];
++  char *p;
++ 
++
+   if (gotoblas) return;
+ 
+-  gotoblas = get_coretype();
++  p = getenv("OPENBLAS_CORETYPE");
++  if ( p )
++  {
++	gotoblas = force_coretype(p);
++  }
++  else
++  {
++  	gotoblas = get_coretype();
++  }
+   
+ #ifdef ARCH_X86
+   if (gotoblas == NULL) gotoblas = &gotoblas_KATMAI;
+@@ -355,6 +429,9 @@ void gotoblas_dynamic_init(void) {
+ #endif
+   
+   if (gotoblas && gotoblas -> init) {
++    strncpy(coren,gotoblas_corename(),20);
++    sprintf(coremsg, "Core: %s\n",coren);
++    openblas_warning(2, coremsg); 
+     gotoblas -> init();
+   } else {
+     openblas_warning(0, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n");
+diff --git a/kernel/arm/gemv_n.c b/kernel/arm/gemv_n.c
+index aedcca9..a295080 100644
+--- a/kernel/arm/gemv_n.c
++++ b/kernel/arm/gemv_n.c
+@@ -61,7 +61,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
+ 		a_ptr += lda;
+ 		ix    += inc_x;
+ 	}
+-
++	return(0);
+ }
+ 	
+ 
+diff --git a/kernel/arm/gemv_t.c b/kernel/arm/gemv_t.c
+index 8fd6a66..f94db40 100644
+--- a/kernel/arm/gemv_t.c
++++ b/kernel/arm/gemv_t.c
+@@ -61,6 +61,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
+ 		iy += inc_y;
+ 		a_ptr += lda;
+ 	}
++	return(0);
+ 
+ }
+