Merge changes Ie115b206,I7ce6cfa2,Iff59b6f4 into amd-staging

* changes:
  Adding __attribute__((constructor/destructor)) for CLANG case.
  Thread Safety: Move bli_init() before and bli_finalize() after main()
  Thread safety: Make the global induced method status array local to thread

frame/base/bli_init.c

@@ -40,8 +40,10 @@ pthread_mutex_t initialize_mutex = PTHREAD_MUTEX_INITIALIZER;
 
 static bool_t bli_is_init = FALSE;
 
-
-err_t bli_init( void )
+// If BLIS is built using a compiler that supports __attribute__((constructor)),
+// then bli_init() will be executed before the application enters main().
+// In that case there is no need to call bli_init() in the application code.
+BLIS_ATTRIB_CTOR err_t bli_init( void )
 {
 	err_t r_val = BLIS_FAILURE;
 
@@ -105,7 +107,10 @@ err_t bli_init( void )
 	return r_val;
 }
 
-err_t bli_finalize( void )
+// If BLIS is built using a compiler that supports __attribute__((destrutor)),
+// then bli_finalize() will be executed after the application exits main().
+// In that case there is no need to call bli_finalize() in the application code.
+BLIS_ATTRIB_DTOR err_t bli_finalize( void )
 {
 	err_t r_val = BLIS_FAILURE;
 

frame/include/bli_macro_defs.h

@@ -64,6 +64,48 @@
 #endif
 
 
+// -- BLIS Thread Local Storage Keyword --
+
+// __thread for TLS is supported by GCC, CLANG, ICC, and IBMC.
+// There is a small risk here as __GNUC__ can also be defined by some other
+// compiler (other than ICC and CLANG which we know define it) that
+// doesn't support __thread, as __GNUC__ is not quite unique to GCC.
+// But the possibility of someone using such non-main-stream compiler
+// for building BLIS is low.
+#if defined(__GNUC__) || defined(__clang__) || defined(__ICC) || defined(__IBMC__)
+  #define BLIS_THREAD_LOCAL __thread
+#else
+  #define BLIS_THREAD_LOCAL
+#endif
+
+
+// -- BLIS constructor/destructor function attribute --
+
+// __attribute__((constructor/destructor)) is supported by GCC only.
+// There is a small risk here as __GNUC__ can also be defined by some other
+// compiler (other than ICC and CLANG which we know define it) that
+// doesn't support this, as __GNUC__ is not quite unique to GCC.
+// But the possibility of someone using such non-main-stream compiler
+// for building BLIS is low.
+
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+  // ICC defines __GNUC__ but doesn't support this
+  #define BLIS_ATTRIB_CTOR
+  #define BLIS_ATTRIB_DTOR
+#elif defined(__clang__)
+  // CLANG supports __attribute__, but its documentation doesn't 
+  // mention support for constructor/destructor. Compiling with 
+  // clang and testing shows that it does support.
+  #define BLIS_ATTRIB_CTOR __attribute__((constructor))
+  #define BLIS_ATTRIB_DTOR __attribute__((destructor))
+#elif defined(__GNUC__)
+  #define BLIS_ATTRIB_CTOR __attribute__((constructor))
+  #define BLIS_ATTRIB_DTOR __attribute__((destructor))
+#else
+  #define BLIS_ATTRIB_CTOR
+  #define BLIS_ATTRIB_DTOR
+#endif
+
 // -- Boolean values --
 
 #ifndef TRUE

frame/ind/bli_l3_ind.c

@@ -60,7 +60,11 @@ static void* bli_l3_ind_oper_fp[BLIS_NUM_IND_METHODS][BLIS_NUM_LEVEL3_OPS] =
 //
 // NOTE: "2" is used instead of BLIS_NUM_FP_TYPES/2.
 //
-static bool_t bli_l3_ind_oper_st[BLIS_NUM_IND_METHODS][BLIS_NUM_LEVEL3_OPS][2] = 
+// BLIS provides APIs to modify this state during runtime. So, one application thread
+// can modify the state, before another starts the corresponding BLIS operation.
+// This is solved by making the induced method status array local to threads.
+
+static BLIS_THREAD_LOCAL bool_t bli_l3_ind_oper_st[BLIS_NUM_IND_METHODS][BLIS_NUM_LEVEL3_OPS][2] = 
 {
         /*   gemm   hemm   herk   her2k  symm   syrk,  syr2k  trmm3  trmm   trsm  */
         /*    c     z    */