Refs #139. Check OS supporting AVX on runtime.

diff --git a/cpuid_x86.c b/cpuid_x86.c
index 79fd20e..ebbbe3f 100644 (file)
--- a/cpuid_x86.c
+++ b/cpuid_x86.c
@@ -114,6 +114,25 @@ static inline int have_excpuid(void){
   return eax & 0xffff;
 }
 
+static inline void xgetbv(int op, int * eax, int * edx){
+  __asm__ __volatile__
+    ("xgetbv": "=a" (*eax), "=d" (*edx) : "c" (op) : "cc");
+}
+
+int support_avx(){
+  int eax, ebx, ecx, edx;
+  int ret=0;
+  
+  cpuid(1, &eax, &ebx, &ecx, &edx);
+  if ((ecx & (1 << 28)) != 0 && (ecx & (1 << 27)) != 0){
+    xgetbv(0, &eax, &edx);
+    if((eax & 6) == 6){
+      ret=1;  //OS support AVX
+    }
+  }
+  return ret;
+}
+
 int get_vendor(void){
   int eax, ebx, ecx, edx;
   char vendor[13];
@@ -195,7 +214,7 @@ int get_cputype(int gettype){
     if ((ecx & (1 << 19)) != 0) feature |= HAVE_SSE4_1;
     if ((ecx & (1 << 20)) != 0) feature |= HAVE_SSE4_2;
 #ifndef NO_AVX
-    if ((ecx & (1 << 28)) != 0) feature |= HAVE_AVX;
+    if (support_avx()) feature |= HAVE_AVX;
 #endif
 
     if (have_excpuid() >= 0x01) {
@@ -991,13 +1010,19 @@ int get_cpuname(void){
          return CPUTYPE_NEHALEM;
        case 10:
          //Intel Core i5-2000 /i7-2000 (Sandy Bridge)
-         return CPUTYPE_SANDYBRIDGE;
+         if(support_avx())
+           return CPUTYPE_SANDYBRIDGE;
+         else
+           return CPUTYPE_NEHALEM; //OS doesn't support AVX
        case 12:
          //Xeon Processor 5600 (Westmere-EP)
          return CPUTYPE_NEHALEM;
        case 13:
          //Intel Core i7-3000 / Xeon E5 (Sandy Bridge)
-         return CPUTYPE_SANDYBRIDGE;
+         if(support_avx())
+           return CPUTYPE_SANDYBRIDGE;
+         else
+           return CPUTYPE_NEHALEM;
        case 15:
          //Xeon Processor E7 (Westmere-EX)
          return CPUTYPE_NEHALEM;
@@ -1006,7 +1031,10 @@ int get_cpuname(void){
       case 3:
        switch (model) {
        case 10:
-         return CPUTYPE_SANDYBRIDGE;
+         if(support_avx())
+           return CPUTYPE_SANDYBRIDGE;
+         else
+           return CPUTYPE_NEHALEM;
        }
        break;
       }
@@ -1350,13 +1378,19 @@ int get_coretype(void){
          return CORE_NEHALEM;
        case 10:
           //Intel Core i5-2000 /i7-2000 (Sandy Bridge)
-          return CORE_SANDYBRIDGE;
+         if(support_avx())
+           return CORE_SANDYBRIDGE;
+         else
+           return CORE_NEHALEM; //OS doesn't support AVX
        case 12:
          //Xeon Processor 5600 (Westmere-EP)
          return CORE_NEHALEM;
        case 13:
           //Intel Core i7-3000 / Xeon E5 (Sandy Bridge)
-          return CORE_SANDYBRIDGE;
+         if(support_avx())
+           return CORE_SANDYBRIDGE;
+         else
+           return CORE_NEHALEM; //OS doesn't support AVX
        case 15:
          //Xeon Processor E7 (Westmere-EX)
          return CORE_NEHALEM;
@@ -1365,7 +1399,10 @@ int get_coretype(void){
       case 3:
        switch (model) {
        case 10:
-         return CORE_SANDYBRIDGE;
+         if(support_avx())
+           return CORE_SANDYBRIDGE;
+         else
+           return CORE_NEHALEM; //OS doesn't support AVX
        }
        break;
       }

diff --git a/driver/others/dynamic.c b/driver/others/dynamic.c
index 45783c5..468ab0d 100644 (file)
--- a/driver/others/dynamic.c
+++ b/driver/others/dynamic.c
@@ -76,6 +76,25 @@ extern gotoblas_t  gotoblas_SANDYBRIDGE;
 
 #define BITMASK(a, b, c) ((((a) >> (b)) & (c)))
 
+static inline void xgetbv(int op, int * eax, int * edx){
+  __asm__ __volatile__
+    ("xgetbv": "=a" (*eax), "=d" (*edx) : "c" (op) : "cc");
+}
+
+int support_avx(){
+  int eax, ebx, ecx, edx;
+  int ret=0;
+  
+  cpuid(1, &eax, &ebx, &ecx, &edx);
+  if ((ecx & (1 << 28)) != 0 && (ecx & (1 << 27)) != 0){
+    xgetbv(0, &eax, &edx);
+    if((eax & 6) == 6){
+      ret=1;  //OS support AVX
+    }
+  }
+  return ret;
+}
+
 static int get_vendor(void){
   int eax, ebx, ecx, edx;
   char vendor[13];
@@ -142,11 +161,25 @@ static gotoblas_t *get_coretype(void){
 
        //Intel Core i5-2000 /i7-2000 (Sandy Bridge)
        //Intel Core i7-3000 / Xeon E5
-       if (model == 10 || model == 13) return &gotoblas_SANDYBRIDGE;
+       if (model == 10 || model == 13) {
+         if(support_avx())
+           return &gotoblas_SANDYBRIDGE;
+         else{
+           fprintf(stderr, "OpenBLAS : Your OS doesn't support AVX. Use Nehalem kernels.\n");
+           return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
+         }
+       }
        return NULL;
       case 3:
        //Intel Sandy Bridge 22nm (Ivy Bridge?)
-       if (model == 10) return &gotoblas_SANDYBRIDGE;
+       if (model == 10) {
+         if(support_avx())
+           return &gotoblas_SANDYBRIDGE;
+         else{
+           fprintf(stderr, "OpenBLAS : Your OS doesn't support AVX. Use Nehalem kernels.\n");
+           return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
+         }
+       }
        return NULL;
       }
       case 0xf:
@@ -239,7 +272,7 @@ void gotoblas_dynamic_init(void) {
   if (gotoblas && gotoblas -> init) {
     gotoblas -> init();
   } else {
-    fprintf(stderr, "GotoBLAS : Architecture Initialization failed. No initialization function found.\n");
+    fprintf(stderr, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n");
     exit(1);
   }