minor corrections to new perf model

src/ctr_comm/seq_tsr.cxx

@@ -245,8 +245,8 @@ double seq_tsr_ctr<dtype>::est_time_fp(int nlyr){
   }
   for (int i=0; i<idx_max; i++){
     if (rev_idx_map[3*i+0] != -1) flops*=edge_len_A[rev_idx_map[3*i+0]];
-    else if (rev_idx_map[3*i+1] != -1) flops*=edge_len_A[rev_idx_map[3*i+1]];
-    else if (rev_idx_map[3*i+2] != -1) flops*=edge_len_A[rev_idx_map[3*i+2]];
+    else if (rev_idx_map[3*i+1] != -1) flops*=edge_len_B[rev_idx_map[3*i+1]];
+    else if (rev_idx_map[3*i+2] != -1) flops*=edge_len_C[rev_idx_map[3*i+2]];
   }
   CTF_free(rev_idx_map);
   return COST_MEMBW*(size_A+size_B+size_C)+COST_FLOP*flops;

src/shared/comm.h

@@ -37,7 +37,7 @@ typedef class CommData {
  
   double estimate_bcast_time(long_int msg_sz) {
 #ifdef BGQ
-    return msg_sz*COST_NETWBW+COST_LATENCY;
+    return msg_sz*(double)COST_NETWBW+COST_LATENCY;
 #else
     return msg_sz*(double)log2(np)*COST_NETWBW;
 #endif
@@ -45,7 +45,7 @@ typedef class CommData {
   
   double estimate_allred_time(long_int msg_sz) {
 #ifdef BGQ
-    return msg_sz*(2.*COST_MEMBW+COST_NETWBW)+COST_LATENCY;
+    return msg_sz*(double)(2.*COST_MEMBW+COST_NETWBW)+COST_LATENCY;
 #else
     return msg_sz*(double)log2(np)*(2.*COST_MEMBW+COST_FLOP+COST_NETWBW);
 #endif