use the query

Signed-off-by: Jeff Hammond <[email protected]>

frame/thread/bli_affinity.c

@@ -43,11 +43,13 @@
 //    https://github.com/flame/blis/issues/604
 //    https://github.com/flame/blis/issues/603 
 
+#include "blis.h"
+
 #ifndef BLIS_ENABLE_AFFINITY
 
 // define the symbol for platforms like Windows and MacOS that do not support the Linux affinity API
 
-int bli_affinity_get_hw_size(int scope)
+dim_t bli_affinity_get_hw_size(bli_affinity_scope_t scope)
 {
     // this is the largest possible value returned by this function
     // and it means that the affinity mask does not constrain the current scope.
@@ -62,13 +64,11 @@ int bli_affinity_get_hw_size(int scope)
 #include <sched.h>
 #include <unistd.h>
 
-#include "blis.h"
-
 // scope is either the calling process or the calling thread:
 //  0 = calling process
 //  1 = calling thread
 
-int bli_affinity_get_hw_size(int scope)
+dim_t bli_affinity_get_hw_size(bli_affinity_scope_t scope)
 {
     int rc;
     int active_cpus;

frame/thread/bli_affinity.h

@@ -35,6 +35,8 @@
 #ifndef BLIS_AFFINITY_H
 #define BLIS_AFFINITY_H
 
-int bli_affinity_get_hw_size(int scope);
+dim_t bli_affinity_get_hw_size(int scope);
 
-#endif BLIS_AFFINITY_H
+typedef enum { process = 0, thread = 1 } bli_affinity_scope_t;
+
+#endif // BLIS_AFFINITY_H

frame/thread/bli_l3_decor_openmp.c

@@ -199,6 +199,12 @@ void bli_l3_thread_decorator_thread_check
      )
 {
 	dim_t n_threads_real = omp_get_num_threads();
+        dim_t n_threads_hwmask;
+        if ( omp_in_parallel() ) {
+            n_threads_hwmask = bli_affinity_get_hw_size(thread);
+        } else {
+            n_threads_hwmask = bli_affinity_get_hw_size(process);
+        }
 
 	// Check if the number of OpenMP threads created within this parallel
 	// region is different from the number of threads that were requested
@@ -241,6 +247,26 @@ void bli_l3_thread_decorator_thread_check
 
 		// Synchronize all threads and continue.
 		_Pragma( "omp barrier" )
+
+                return;
+	}
+
+	// Check if the number of OpenMP threads created within this parallel
+	// region is different from the number of threads that are available
+	// to BLIS in the calling context.
+	if ( n_threads_hwsize < n_threads || n_threads_hwsize < n_threads_real)
+	{
+                bli_print_msg( "The affinity mask on this process does not have "
+                               "enough HW threads for your requested SW threads.",
+                               __FILE__, __LINE__ );
+
+		bli_thrcomm_init( n_threads_hwsize, gl_comm );
+		bli_rntm_set_num_threads_only( n_threads_hwsize, rntm );
+#warning HELP ME HERE
+		bli_rntm_set_ways_only( 1, 1, 1, 1, 1, rntm );
+
+		// Synchronize all threads and continue.
+		_Pragma( "omp barrier" )
 	}
 }