BLIS: Nested Parallelism issues (3)

Bugfix for parallel BLAS1 and BLAS2 routines. Threading information
was not being set correctly when initializing local rntm from global.

Also ensure th_rntm is initialized along with global_rntm by updating
it in bli_thread_init(), called by bli_init_once()

AMD-Internal: [CPUPL-2433]
Change-Id: Iba658f87ae13fe16a57ca1fc279e149b7fa294cf

frame/base/bli_rntm.c

@@ -54,9 +54,12 @@ void bli_rntm_init_from_global( rntm_t* rntm )
 
 	dim_t jc, pc, ic, jr, ir;
 
-	// We must ensure that global_rntm and tl_rntm have been initialized.
+	// We must ensure that global_rntm has been initialized
 	bli_init_once();
 
+	// We must also ensure that tl_rntm has been updated.
+	bli_thread_update_tl();
+
 	// Acquire the mutex protecting global_rntm.
 	bli_pthread_mutex_lock( &global_rntm_mutex );
 

frame/thread/bli_thread.c

@@ -64,6 +64,8 @@ void bli_thread_init( void )
 	// to ensure all thread local get information from any BLIS environment
 	// variables set, as these are not re-read for performance reasons.
 	bli_thread_init_rntm_from_env( &global_rntm );
+	// Initialize tl_rntm.
+	bli_thread_update_rntm_from_env( &tl_rntm );
 }
 
 void bli_thread_update_tl( void )