This patch enables the use of KMP_AFFINITY=balanced on non-MIC Architectures. The restriction for using balanced affinity on non-MIC architectures is it only works for one-package machines.

git-svn-id: https://llvm.org/svn/llvm-project/openmp/trunk@225794 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 19 insertions, 39 deletions

diff --git a/runtime/src/kmp.h b/runtime/src/kmp.h
index 448ee87..83ecb15 100644
--- a/runtime/src/kmp.h
+++ b/runtime/src/kmp.h
@@ -679,9 +679,7 @@ enum affinity_type {
     affinity_compact,
     affinity_scatter,
     affinity_explicit,
-#if KMP_MIC
     affinity_balanced,
-#endif
     affinity_disabled,  // not used outsize the env var parser
     affinity_default
 };
@@ -2336,9 +2334,7 @@ typedef struct KMP_ALIGN_CACHE kmp_base_team {
     int                      t_first_place;  // first & last place in parent thread's partition.
     int                      t_last_place;   // Restore these values to master after par region.
 #endif // OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
-#if KMP_MIC
     int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via omp_set_num_threads() call
-#endif
 
     // Read/write by workers as well -----------------------------------------------------------------------
 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
diff --git a/runtime/src/kmp_affinity.cpp b/runtime/src/kmp_affinity.cpp
index d6821e0..a9338c2 100644
--- a/runtime/src/kmp_affinity.cpp
+++ b/runtime/src/kmp_affinity.cpp
@@ -3819,9 +3819,8 @@ __kmp_aux_affinity_initialize(void)
         }
         goto sortAddresses;
 
-# if KMP_MIC
         case affinity_balanced:
-        // Balanced works only for the case of a single package and uniform topology
+        // Balanced works only for the case of a single package
         if( nPackages > 1 ) {
             if( __kmp_affinity_verbose || __kmp_affinity_warnings ) {
                 KMP_WARNING( AffBalancedNotAvail, "KMP_AFFINITY" );
@@ -3872,7 +3871,6 @@ __kmp_aux_affinity_initialize(void)
 
             break;
         }
-# endif
 
         sortAddresses:
         //
@@ -4019,10 +4017,7 @@ __kmp_affinity_set_init_mask(int gtid, int isa_root)
     if (__kmp_nested_proc_bind.bind_types[0] == proc_bind_intel)
 # endif
     {
-        if ((__kmp_affinity_type == affinity_none)
-# if KMP_MIC
-          || (__kmp_affinity_type == affinity_balanced)
-# endif
+        if ((__kmp_affinity_type == affinity_none) || (__kmp_affinity_type == affinity_balanced)
           ) {
 # if KMP_OS_WINDOWS && KMP_ARCH_X86_64
             if (__kmp_num_proc_groups > 1) {
@@ -4397,7 +4392,6 @@ __kmp_aux_get_affinity_mask_proc(int proc, void **mask)
     return KMP_CPU_ISSET(proc, (kmp_affin_mask_t *)(*mask));
 }
 
-# if KMP_MIC
 
 // Dynamic affinity settings - Affinity balanced
 void __kmp_balanced_affinity( int tid, int nthreads )
@@ -4626,8 +4620,6 @@ void __kmp_balanced_affinity( int tid, int nthreads )
     }
 }
 
-# endif /* KMP_MIC */
-
 #else
     // affinity not supported
 
diff --git a/runtime/src/kmp_barrier.cpp b/runtime/src/kmp_barrier.cpp
index 79cfd3a..dc1d0ec 100644
--- a/runtime/src/kmp_barrier.cpp
+++ b/runtime/src/kmp_barrier.cpp
@@ -1536,12 +1536,12 @@ __kmp_fork_barrier(int gtid, int tid)
     kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
     if (proc_bind == proc_bind_intel) {
 #endif
-#if KMP_MIC
+#if KMP_AFFINITY_SUPPORTED
         // Call dynamic affinity settings
         if(__kmp_affinity_type == affinity_balanced && team->t.t_size_changed) {
             __kmp_balanced_affinity(tid, team->t.t_nproc);
         }
-#endif
+#endif // KMP_AFFINITY_SUPPORTED
 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
     }
     else if ((proc_bind != proc_bind_false)
diff --git a/runtime/src/kmp_runtime.c b/runtime/src/kmp_runtime.c
index 30c1461..e7106aa 100644
--- a/runtime/src/kmp_runtime.c
+++ b/runtime/src/kmp_runtime.c
@@ -2268,10 +2268,8 @@ __kmp_set_num_threads( int new_nth, int gtid )
             KMP_DEBUG_ASSERT( hot_team->t.t_threads[f] != NULL );
             hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
         }
-#if KMP_MIC
         // Special flag in case omp_set_num_threads() call
         hot_team->t.t_size_changed = -1;
-#endif
     }
 
 }
@@ -2837,9 +2835,7 @@ __kmp_initialize_root( kmp_root_t *root )
     // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
     hot_team->t.t_sched.r_sched_type = r_sched.r_sched_type;
     hot_team->t.t_sched.chunk        = r_sched.chunk;
-#if KMP_MIC
     hot_team->t.t_size_changed = 0;
-#endif
 
 }
 
@@ -4411,7 +4407,6 @@ __kmp_allocate_team( kmp_root_t *root, int new_nproc, int max_nproc,
            put that case first. */
         if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
             KA_TRACE( 20, ("__kmp_allocate_team: reusing hot team\n" ));
-#if KMP_MIC
             // This case can mean that omp_set_num_threads() was called and the hot team size
             // was already reduced, so we check the special flag
             if ( team->t.t_size_changed == -1 ) {
@@ -4419,7 +4414,6 @@ __kmp_allocate_team( kmp_root_t *root, int new_nproc, int max_nproc,
             } else {
                 team->t.t_size_changed = 0;
             }
-#endif
 
             // TODO???: team->t.t_max_active_levels = new_max_active_levels;
             team->t.t_sched =  new_icvs->sched;
@@ -4451,9 +4445,7 @@ __kmp_allocate_team( kmp_root_t *root, int new_nproc, int max_nproc,
         else if( team->t.t_nproc > new_nproc ) {
             KA_TRACE( 20, ("__kmp_allocate_team: decreasing hot team thread count to %d\n", new_nproc ));
 
-#if KMP_MIC
             team->t.t_size_changed = 1;
-#endif
             if ( __kmp_tasking_mode != tskm_immediate_exec ) {
                 kmp_task_team_t *task_team = team->t.t_task_team;
                 if ( ( task_team != NULL ) && TCR_SYNC_4(task_team->tt.tt_active) ) {
@@ -4541,9 +4533,7 @@ __kmp_allocate_team( kmp_root_t *root, int new_nproc, int max_nproc,
 
             KA_TRACE( 20, ("__kmp_allocate_team: increasing hot team thread count to %d\n", new_nproc ));
 
-#if KMP_MIC
             team->t.t_size_changed = 1;
-#endif
 
 
 #if KMP_NESTED_HOT_TEAMS
diff --git a/runtime/src/kmp_settings.c b/runtime/src/kmp_settings.c
index 016e5a3..53ab1d7 100644
--- a/runtime/src/kmp_settings.c
+++ b/runtime/src/kmp_settings.c
@@ -2128,14 +2128,12 @@ __kmp_parse_affinity_env( char const * name, char const * value,
             __kmp_nested_proc_bind.bind_types[0] = proc_bind_intel;
 # endif
             buf = next;
-# if KMP_MIC
         } else if (__kmp_match_str("balanced", buf, (const char **)&next)) {
             set_type( affinity_balanced );
 #  if OMP_40_ENABLED
             __kmp_nested_proc_bind.bind_types[0] = proc_bind_intel;
 #  endif
             buf = next;
-# endif
         } else if (__kmp_match_str("disabled", buf, (const char **)&next)) {
             set_type( affinity_disabled );
 # if OMP_40_ENABLED
@@ -2309,7 +2307,6 @@ __kmp_parse_affinity_env( char const * name, char const * value,
                 KMP_WARNING( AffManyParamsForLogic, name, number[ 1 ] );
             }; // if
         } break;
-# if KMP_MIC
         case affinity_balanced: {
             if ( count > 0 ) {
                 *out_compact = number[ 0 ];
@@ -2318,16 +2315,20 @@ __kmp_parse_affinity_env( char const * name, char const * value,
                 *out_offset = number[ 1 ];
             }; // if
 
-            // If granularity is neither thread nor core let it be default value=fine
-            if( __kmp_affinity_gran != affinity_gran_default && __kmp_affinity_gran != affinity_gran_fine
-                && __kmp_affinity_gran != affinity_gran_thread && __kmp_affinity_gran != affinity_gran_core ) {
+            if ( __kmp_affinity_gran == affinity_gran_default ) {
+# if KMP_MIC
                 if( __kmp_affinity_verbose || __kmp_affinity_warnings ) {
-                    KMP_WARNING( AffGranUsing, "KMP_AFFINITY", "core" );
+                    KMP_WARNING( AffGranUsing, "KMP_AFFINITY", "fine" );
                 }
                 __kmp_affinity_gran = affinity_gran_fine;
+# else
+                if( __kmp_affinity_verbose || __kmp_affinity_warnings ) {
+                    KMP_WARNING( AffGranUsing, "KMP_AFFINITY", "core" );
+                }
+                __kmp_affinity_gran = affinity_gran_core;
+# endif /* KMP_MIC */
             }
         } break;
-# endif
         case affinity_scatter:
         case affinity_compact: {
             if ( count > 0 ) {
@@ -2468,12 +2469,10 @@ __kmp_stg_print_affinity( kmp_str_buf_t * buffer, char const * name, void * data
             __kmp_str_buf_print( buffer, "%s=[%s],%s", "proclist",
               __kmp_affinity_proclist, "explicit" );
             break;
-# if KMP_MIC
         case affinity_balanced:
             __kmp_str_buf_print( buffer, "%s,%d,%d", "balanced",
               __kmp_affinity_compact, __kmp_affinity_offset );
             break;
-# endif
         case affinity_disabled:
             __kmp_str_buf_print( buffer, "%s", "disabled");
             break;
@@ -3026,10 +3025,15 @@ __kmp_stg_parse_proc_bind( char const * name, char const * value, void * data )
     __kmp_stg_parse_bool( name, value, & enabled );
     if ( enabled ) {
             //
-            // OMP_PROC_BIND => granularity=core,scatter
+            // OMP_PROC_BIND => granularity=fine,scatter on MIC
+            // OMP_PROC_BIND => granularity=core,scatter elsewhere
             //
             __kmp_affinity_type = affinity_scatter;
+#  if KMP_MIC
+            __kmp_affinity_gran = affinity_gran_fine;
+#  else
             __kmp_affinity_gran = affinity_gran_core;
+#  endif /* KMP_MIC */
     }
     else {
         __kmp_affinity_type = affinity_none;
@@ -4987,9 +4991,7 @@ __kmp_env_initialize( char const * string ) {
           && ( FIND( aff_str, "compact" ) == NULL )
           && ( FIND( aff_str, "scatter" ) == NULL )
           && ( FIND( aff_str, "explicit" ) == NULL )
-# if KMP_MIC
           && ( FIND( aff_str, "balanced" ) == NULL )
-# endif
           && ( FIND( aff_str, "disabled" ) == NULL ) ) {
             __kmp_affinity_notype = __kmp_stg_find( "KMP_AFFINITY"  );
         }