Merge pull request #23 from davidozog/pr/SOS_v1.5.1_sync

Sync with SOS v1.5.1

test/apps/gups.c

@@ -322,25 +322,11 @@ SHMEMRandomAccess(void)
   uint64_t NumUpdates; /* total number of updates to table */
   uint64_t ProcNumUpdates; /* number of updates per processor */
 
-  static long pSync_bcast[SHMEM_BCAST_SYNC_SIZE];
-  static long long int llpWrk[SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-
-  static long pSync_reduce[SHMEM_REDUCE_SYNC_SIZE];
-  static int ipWrk[SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-
   FILE *outFile = NULL;
   double *GUPs;
   double *temp_GUPs;
 
 
-  for (i = 0; i < SHMEM_BCAST_SYNC_SIZE; i += 1){
-        pSync_bcast[i] = SHMEM_SYNC_VALUE;
-  }
-
-  for (i = 0; i < SHMEM_REDUCE_SYNC_SIZE; i += 1){
-        pSync_reduce[i] = SHMEM_SYNC_VALUE;
-  }
-
   SHMEMGUPs = -1;
   GUPs = &SHMEMGUPs;
 
@@ -399,7 +385,7 @@ SHMEMRandomAccess(void)
   if (! HPCC_PELock) sAbort = 1;
 
   shmem_barrier_all();
-  shmem_int_sum_to_all(&rAbort, &sAbort, 1, 0, 0, NumProcs, ipWrk, pSync_reduce);
+  shmem_int_sum_reduce(SHMEM_TEAM_WORLD, &rAbort, &sAbort, 1);
   shmem_barrier_all();
 
   if (rAbort > 0) {
@@ -468,7 +454,7 @@ SHMEMRandomAccess(void)
   /* distribute result to all nodes */
   temp_GUPs = GUPs;
   shmem_barrier_all();
-  shmem_broadcast64(GUPs,temp_GUPs,1,0,0,0,NumProcs,pSync_bcast);
+  shmem_double_broadcast(SHMEM_TEAM_WORLD,GUPs,temp_GUPs,1,0);
   shmem_barrier_all();
 
   /* Verification phase */
@@ -494,7 +480,7 @@ SHMEMRandomAccess(void)
   }
 
   shmem_barrier_all();
-  shmem_longlong_sum_to_all( (long long *)&GlbNumErrors,  (long long *)&NumErrors, 1, 0,0, NumProcs,llpWrk, pSync_reduce);
+  shmem_longlong_sum_reduce(SHMEM_TEAM_WORLD, (long long *)&GlbNumErrors,  (long long *)&NumErrors, 1);
   shmem_barrier_all();
 
   /* End timed section */

test/performance/shmem_perf_suite/bi_dir.h

@@ -75,13 +75,13 @@ static inline void bi_bw_put(int len, perf_metrics_t * const metric_info)
     }
 
     if (streaming_node(metric_info)) {
-        shmem_int_p(&fin, 1, dest);
+        shmem_int_atomic_set(&fin, 1, dest);
         shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 0);
         end = perf_shmemx_wtime();
         calc_and_print_results(end, start, len, metric_info);
     } else {
         shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 1);
-        shmem_int_p(&fin, 0, dest);
+        shmem_int_atomic_set(&fin, 0, dest);
     }
 
 }
@@ -146,13 +146,13 @@ static inline void bi_bw_get(int len, perf_metrics_t * const metric_info)
     } 
 
     if (streaming_node(metric_info)) {
-        shmem_int_p(&fin, 1, dest);
+        shmem_int_atomic_set(&fin, 1, dest);
         shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 0);
         end = perf_shmemx_wtime();
         calc_and_print_results(end, start, len, metric_info);
     } else {
         shmem_int_wait_until(&fin, SHMEM_CMP_EQ, 1);
-        shmem_int_p(&fin, 0, dest);
+        shmem_int_atomic_set(&fin, 0, dest);
     }
 }
 

test/performance/shmem_perf_suite/bw_common.h

@@ -141,17 +141,18 @@ void print_data_results(double bw, double mr, const perf_metrics_t * const data,
 static inline
 void calc_and_print_results(double end_t, double start_t, int len,
                             perf_metrics_t * const metric_info) {
-    int stride = 0, start_pe = 0, nPEs = 0;
+    int start_pe = 0, nPEs = metric_info->num_pes;
     static double pe_bw_sum, bw = 0.0; /*must be symmetric for reduction*/
     double pe_bw_avg = 0.0, pe_mr_avg = 0.0;
     int nred_elements = 1;
-    static double pwrk[SHMEM_REDUCE_MIN_WRKDATA_SIZE];
     static double pe_time_start, pe_time_end,
                   end_time_max = 0.0, start_time_min = 0.0;
     double total_t = 0.0, total_t_max = 0.0;
     int multiplier = 1;
+    shmem_team_t sync_team;
 
-    PE_set_used_adjustments(&nPEs, &stride, &start_pe, metric_info);
+    PE_set_used_adjustments(&nPEs, &start_pe, metric_info);
+    sync_team = (start_pe == 0) ? streaming_team : target_team;
 
     /* 2x as many messages at once for bi-directional */
     if(metric_info->b_type == BI_DIR)
@@ -188,16 +189,12 @@ void calc_and_print_results(double end_t, double start_t, int len,
 
     pe_time_start = start_t;
     pe_time_end = end_t;
-    shmem_barrier(start_pe, stride, nPEs, bar_psync);
+    shmem_team_sync(sync_team);
     if (metric_info->cstyle != COMM_INCAST) {
         if (nPEs >= 2) {
-            shmem_double_min_to_all(&start_time_min, &pe_time_start, nred_elements,
-                                start_pe, stride, nPEs, pwrk,
-                                red_psync);
-            shmem_barrier(start_pe, stride, nPEs, bar_psync);
-            shmem_double_max_to_all(&end_time_max, &pe_time_end, nred_elements,
-                                start_pe, stride, nPEs, pwrk,
-                                red_psync);
+            shmem_double_min_reduce(streaming_team, &start_time_min, &pe_time_start, nred_elements);
+            shmem_team_sync(sync_team);
+            shmem_double_max_reduce(streaming_team, &end_time_max, &pe_time_end, nred_elements);
         } else if (nPEs == 1) {
             start_time_min = pe_time_start;
             end_time_max = pe_time_end;
@@ -226,9 +223,7 @@ void calc_and_print_results(double end_t, double start_t, int len,
         pe_bw_sum = bw;
     } else {
         if (nPEs >= 2) {
-            shmem_double_sum_to_all(&pe_bw_sum, &bw, nred_elements,
-                                start_pe, stride, nPEs, pwrk,
-                                red_psync);
+            shmem_double_sum_reduce(streaming_team, &pe_bw_sum, &bw, nred_elements);
         } else if (nPEs == 1) {
             pe_bw_sum = bw;
         }
@@ -421,9 +416,8 @@ int bw_init_data_stream(perf_metrics_t * const metric_info,
 #if defined(ENABLE_THREADS)
     thread_safety_validation_check(metric_info);
 #endif
-    init_psync_arrays();
 
-    if(only_even_PEs_check(metric_info->my_node, metric_info->num_pes) != 0) {
+    if (only_even_PEs_check(metric_info->my_node, metric_info->num_pes) != 0) {
         return -1;
     }
 
@@ -433,6 +427,10 @@ int bw_init_data_stream(perf_metrics_t * const metric_info,
     metric_info->dest = aligned_buffer_alloc(metric_info->max_len * metric_info->nthreads);
     init_array(metric_info->dest, metric_info->max_len * metric_info->nthreads, metric_info->my_node);
 
+    if (create_teams(metric_info) != 0) {
+        return -1;
+    } 
+
     return 0;
 }
 

test/performance/shmem_perf_suite/common.h

@@ -168,17 +168,15 @@ typedef struct perf_metrics {
     int individual_report;
 } perf_metrics_t;
 
-/* psync arrays used in metric calculation */
-long red_psync[SHMEM_REDUCE_SYNC_SIZE];
-long bar_psync[SHMEM_BARRIER_SYNC_SIZE];
+
+shmem_team_t streaming_team, target_team;
 
 /* default settings with no input provided */
 static inline
 void set_metric_defaults(perf_metrics_t *metric_info) {
     char *val = NULL;
     metric_info->trials_multiplier = 1.0; /* Default 1 */
     val = getenv("SHMEM_PERF_SUITE_TRIALS_MULTIPLIER");
-
     if (val && strlen(val))
         metric_info->trials_multiplier = atof(val);
 
@@ -220,17 +218,6 @@ void update_metrics(perf_metrics_t *metric_info) {
     metric_info->midpt = metric_info->num_pes / 2;
 }
 
-/* init psync arrays */
-static inline
-void init_psync_arrays(void) {
-    int i;
-    for(i = 0; i < SHMEM_REDUCE_SYNC_SIZE; i++)
-        red_psync[i] = SHMEM_SYNC_VALUE;
-
-    for(i = 0; i < SHMEM_BARRIER_SYNC_SIZE; i++)
-        bar_psync[i] = SHMEM_SYNC_VALUE;
-}
-
 /* return microseconds */
 double perf_shmemx_wtime(void);
 
@@ -572,6 +559,7 @@ void thread_safety_validation_check(perf_metrics_t * const metric_info) {
 }
 #endif
 
+/* Only even number of PEs are allowed for performance tests */
 static inline
 int only_even_PEs_check(int my_node, int num_pes) {
     if (num_pes % 2 != 0) {
@@ -612,7 +600,7 @@ int partner_node(const perf_metrics_t * const my_info)
 static inline
 int streaming_node(const perf_metrics_t * const my_info)
 {
-    if(my_info->cstyle == COMM_PAIRWISE) {
+    if (my_info->cstyle == COMM_PAIRWISE) {
         return (my_info->my_node < my_info->szinitiator);
     } else {
         assert(my_info->cstyle == COMM_INCAST);
@@ -643,16 +631,9 @@ int check_hostname_validation(const perf_metrics_t * const my_info) {
 
     int hostname_status = -1;
 
-    /* hostname_size should be a length divisible by 4 */
-    int hostname_size = (MAX_HOSTNAME_LEN % 4 == 0) ? MAX_HOSTNAME_LEN :
-                         MAX_HOSTNAME_LEN + (4 - MAX_HOSTNAME_LEN % 4);
+    int hostname_size = MAX_HOSTNAME_LEN;
     int i, errors = 0;
 
-    /* pSync for fcollect of hostnames */
-    static long pSync_collect[SHMEM_COLLECT_SYNC_SIZE];
-    for (i = 0; i < SHMEM_COLLECT_SYNC_SIZE; i++)
-        pSync_collect[i] = SHMEM_SYNC_VALUE;
-
     char *hostname = (char *) shmem_malloc (hostname_size * sizeof(char));
     char *dest = (char *) shmem_malloc (my_info->num_pes * hostname_size *
                                         sizeof(char));
@@ -669,9 +650,7 @@ int check_hostname_validation(const perf_metrics_t * const my_info) {
     }
     shmem_barrier_all();
 
-    /* nelems needs to be updated based on 32-bit API */
-    shmem_fcollect32(dest, hostname, hostname_size/4, 0, 0, my_info->num_pes,
-                     pSync_collect);
+    shmem_char_fcollect(SHMEM_TEAM_WORLD, dest, hostname, hostname_size);
 
     char *snode_name = NULL;
     char *tnode_name = NULL;
@@ -769,16 +748,16 @@ void large_message_metric_chg(perf_metrics_t * const metric_info, int len) {
 static inline
 red_PE_set validation_set(perf_metrics_t * const my_info, int *nPEs)
 {
-    if(my_info->cstyle == COMM_PAIRWISE) {
-        if(streaming_node(my_info)) {
+    if (my_info->cstyle == COMM_PAIRWISE) {
+        if (streaming_node(my_info)) {
             *nPEs = my_info->szinitiator;
             return FIRST_HALF;
-        } else if(target_node(my_info)) {
+        } else if (target_node(my_info)) {
             *nPEs = my_info->sztarget;
             return SECOND_HALF;
         } else {
             fprintf(stderr, "Warning: you are getting data from a node that "
-                "wasn't a part of the perf set \n ");
+                            "wasn't a part of the perf set \n ");
             return 0;
         }
     } else {
@@ -788,11 +767,8 @@ red_PE_set validation_set(perf_metrics_t * const my_info, int *nPEs)
     }
 }
 
-/* reduction to collect performance results from PE set
- * then start_pe will print results --- assumes num_pes is even */
 static inline
-void PE_set_used_adjustments(int *nPEs, int *stride, int *start_pe,
-                             perf_metrics_t * const my_info) {
+void PE_set_used_adjustments(int *nPEs, int *start_pe, perf_metrics_t * const my_info) {
     red_PE_set PE_set = validation_set(my_info, nPEs);
 
     if(PE_set == FIRST_HALF || PE_set == FULL_SET) {
@@ -802,8 +778,6 @@ void PE_set_used_adjustments(int *nPEs, int *stride, int *start_pe,
         assert(PE_set == SECOND_HALF);
         *start_pe = my_info->midpt;
     }
-
-    *stride = 0; /* back to back PEs */
 }
 
 static
@@ -822,3 +796,41 @@ void print_header(perf_metrics_t * const metric_info) {
 #endif
     printf("\n");
 }
+
+static
+int create_streaming_team(perf_metrics_t * const metric_info) {
+    shmem_team_split_strided(SHMEM_TEAM_WORLD, 0, 1, metric_info->num_pes / 2, NULL, 0, &streaming_team);
+
+    int my_pe = metric_info->my_node;
+    if (streaming_team == SHMEM_TEAM_INVALID && (my_pe >= 0 && my_pe < metric_info->num_pes / 2)) {
+        fprintf(stderr, "PE %d: Streaming team creation failed\n", metric_info->my_node);
+        return -1;
+    }
+
+    return 0;
+}
+
+static
+int create_target_team(perf_metrics_t * const metric_info) {
+    shmem_team_split_strided(SHMEM_TEAM_WORLD, metric_info->midpt, 1, metric_info->num_pes / 2, NULL, 0, &target_team);
+
+    int my_pe = metric_info->my_node;
+    if (target_team == SHMEM_TEAM_INVALID && (my_pe >= metric_info->midpt && my_pe < metric_info->num_pes)) {
+        fprintf(stderr, "PE %d: Target team creation failed\n", metric_info->my_node);
+        return -1;
+    }
+
+    return 0;
+}
+
+/* Create two teams: streaming and target. 
+ * PEs [0, 1, ..., npes/2-1] will be in streaming_team and
+ * PEs [npes/2, npes/2+1, ..., npes-1] in target_team.  */ 
+static 
+int create_teams(perf_metrics_t * const metric_info) {
+    int ret = create_streaming_team(metric_info);
+    if (!ret)
+        return create_target_team(metric_info);
+
+    return ret;
+}

test/performance/shmem_perf_suite/latency_common.h

@@ -51,13 +51,14 @@ void print_latency_header(void) {
 static inline 
 void calc_and_print_results(double start, double end, int len,
                             perf_metrics_t * const metric_info) {
-    int stride = 0, start_pe = 0, nPEs = 0;
+    int start_pe = 0, nPEs = metric_info->num_pes;
     int nred_elements = 1;
     static double latency = 0.0, avg_latency = 0.0;
-    static double pwrk[SHMEM_REDUCE_MIN_WRKDATA_SIZE];
-
-    PE_set_used_adjustments(&nPEs, &stride, &start_pe, metric_info);
-
+    shmem_team_t sync_team;
+
+    PE_set_used_adjustments(&nPEs, &start_pe, metric_info);
+    sync_team = (start_pe == 0) ? streaming_team : target_team;
+
     if (end > 0 && start > 0 && (end - start) > 0) {
         latency = (end - start) / metric_info->trials;
     } else {
@@ -69,12 +70,10 @@ void calc_and_print_results(double start, double end, int len,
         printf("Individual latency for PE %6d is %10.2f\n",
                 metric_info->my_node, latency);
     }
-    shmem_barrier(start_pe, stride, nPEs, bar_psync);
+    shmem_team_sync(sync_team);
 
     if (nPEs >= 2) {
-        shmem_double_sum_to_all(&avg_latency, &latency, 
-                                nred_elements, start_pe, stride,
-                                nPEs, pwrk, red_psync);
+        shmem_double_sum_reduce(streaming_team, &avg_latency, &latency, nred_elements); 
         avg_latency /= nPEs;
     } else {
         avg_latency = latency;
@@ -175,7 +174,6 @@ int latency_init_resources(int argc, char *argv[],
 #if defined(ENABLE_THREADS)
     thread_safety_validation_check(metric_info);
 #endif
-    init_psync_arrays();
 
     if(only_even_PEs_check(metric_info->my_node, metric_info->num_pes) != 0) {
         return -1;
@@ -193,6 +191,10 @@ int latency_init_resources(int argc, char *argv[],
     metric_info->target = shmalloc(sizeof(long));
 #endif
 
+    if (create_teams(metric_info) != 0) {
+        return -1;
+    }
+
     return 0;
 }