Changeset 823 for soft/giet_vm/applications/rosenfeld/src-par/mca_rosenfeld.c

Timestamp:

Jun 14, 2016, 5:23:56 PM (8 years ago)

Author:

meunier

Message:

• Improved scripts for simulations and graphes
• Continued to clean up the lib nrc2 (from nrio2x.x to nrmem1.c)
• Added a version (Fast - Parmerge - No stats)

File:

• soft/giet_vm/applications/rosenfeld/src-par/mca_rosenfeld.c (modified) (8 diffs)

soft/giet_vm/applications/rosenfeld/src-par/mca_rosenfeld.c

@@ -142 +142 @@
 }
 #endif // FEATURES && !PARMERGE
+
+
+#if !FEATURES && PARMERGE
+// -----------------------------------------------------------------------------------------------------------
+static bool SetRoot_Parallel_Rosenfeld_Dist(uint32 ** D, uint32 root, uint32 eps, int shift, RegionStats ** F)
+// -----------------------------------------------------------------------------------------------------------
+{
+    assert(root != 0 && eps != 0);
+
+    MCA_VERBOSE3(printf("F(%d) += F(%d)\n", eps, root));
+    
+    int mask = (1 << shift) - 1;
+
+    uint32 r1 = root >> shift;
+    uint32 r0 = root & mask;
+    
+    uint32 e1 = eps >> shift;
+    uint32 e0 = eps & mask;
+
+    // Locking towards the root (first root, then eps)
+    pthread_spin_lock(&F[r1][r0].lock);
+    pthread_spin_lock(&F[e1][e0].lock);
+    if (D[e1][e0] != eps || D[r1][r0] != root) {
+        // Someone change the root of epsilon or "root", need to find the new root
+        pthread_spin_unlock(&F[e1][e0].lock);
+        pthread_spin_unlock(&F[r1][r0].lock);
+        return false;
+    }
+
+    D[r1][r0] = eps;
+    
+    pthread_spin_unlock(&F[e1][e0].lock);
+    pthread_spin_unlock(&F[r1][r0].lock);
+    return true;
+}
+#endif // !FEATURES && PARMERGE
 
 
@@ -585 +621 @@
 
 
-#if FAST && FEATURES && PARMERGE && !ARSP
+#if FAST && PARMERGE && !ARSP // Valid for FEATURES and !FEATURES
 // ---------------------------------------------------------------------------------------------------------------------------
-static void vuse2_Parallel_Features_Rosenfeld_Dist(uint32 ed, uint32 el, uint32 * T, uint32 ** D, int alpha, RegionStats ** F)
+static void vuse2_Parallel_Rosenfeld_Dist(uint32 ed, uint32 el, uint32 * T, uint32 ** D, int alpha, RegionStats ** F)
 // ---------------------------------------------------------------------------------------------------------------------------
 {
@@ -610 +646 @@
         // qui a fait un test
         if (rd < rl) {
-            ok = SetRoot_Parallel_Features_Rosenfeld_Dist(D, rl, rd, alpha, F);
+            // Features or No Features depending on config
+            ok = SetRoot_Parallel_FNF(D, rl, rd, alpha, F);
         }
         else {
-            ok = SetRoot_Parallel_Features_Rosenfeld_Dist(D, rd, rl, alpha, F);
+            ok = SetRoot_Parallel_FNF(D, rd, rl, alpha, F);
         }
     } while (!ok);
 }
 
-// FAST && FEATURES && PARMERGE && !ARSP
+// FAST && PARMERGE && !ARSP
 
 // -----------------------------------------------------------------------------------------------------------------------------------------
-static void vuse3_Parallel_Features_Rosenfeld_Dist(uint32 ed1, uint32 ed2, uint32 el3, uint32 * T, uint32 ** D, int alpha, RegionStats ** F)
+static void vuse3_Parallel_Rosenfeld_Dist(uint32 ed1, uint32 ed2, uint32 el3, uint32 * T, uint32 ** D, int alpha, RegionStats ** F)
 // -----------------------------------------------------------------------------------------------------------------------------------------
 {
@@ -649 +686 @@
         ok3 = true;
         if (r1 > eps) {
-            ok1 = SetRoot_Parallel_Features_Rosenfeld_Dist(D, r1, eps, alpha, F);
+            ok1 = SetRoot_Parallel_FNF(D, r1, eps, alpha, F);
         }
         if (r2 > eps && r2 != r1) {
-            ok2 = SetRoot_Parallel_Features_Rosenfeld_Dist(D, r2, eps, alpha, F);
+            ok2 = SetRoot_Parallel_FNF(D, r2, eps, alpha, F);
         }
         if (r3 > eps && r3 != r2 && r3 != r1) {
-            ok3 = SetRoot_Parallel_Features_Rosenfeld_Dist(D, r3, eps, alpha, F);
+            ok3 = SetRoot_Parallel_FNF(D, r3, eps, alpha, F);
         }
     } while (!(ok1 && ok2 && ok3));
 }
-#endif // FAST && FEATURES && PARMERGE && !ARSP
+#endif // FAST && PARMERGE && !ARSP
 
 
@@ -1381 +1418 @@
     uint32 e0 = mca->e0;
     uint32 e1 = mca->e1;
+    uint32 ne_prev = mca->ne_prev;
     uint32 ne = e0 - 1;
     uint32 nr = 0;
@@ -1390 +1428 @@
     RegionStats * stats = mca->stats;
 
-    // reset sous optimal (pour le moment = voir region32)
-    if (mca->p == 0) {
-        set_ui32vector_j(T, e0 - 1, e1); // car e0 = 1, on a besoin que T[0] = 0 pour FindRoot
+    CLOCK_THREAD_START_STEP(mca->p, 0);
+
+    set_ui32vector_j(T, e0, ne_prev);
 #if FEATURES
-        zero_RegionStatsVector(stats, e0 - 1, e1);
+    zero_RegionStatsVector(stats, e0, ne_prev);
 #endif
-    }
-    else {
-        set_ui32vector_j(T, e0, e1);
-#if FEATURES
-        zero_RegionStatsVector(stats, e0, e1);
-#endif
-    }
 
     if (mca->p == 0) {
@@ -1411 +1442 @@
     // -- Etiquetage d'une bande -- //
     // ---------------------------- //
-
-    CLOCK_THREAD_START_STEP(mca->p, 0);
 
     ne = line0Labeling_Rosenfeld(X, i0, width, E, T, ne);
@@ -1644 +1673 @@
 
     CLOCK_THREAD_START(mca->p);
-    CLOCK_THREAD_COMPUTE_START(mca->p);
-
-    MCA_Scatter_ImageX(mca);
-    pthread_barrier_wait(&main_barrier);
-
-    MCA_Label_Rosenfeld_PAR1(mca);
-    pthread_barrier_wait(&main_barrier);
-   
+
+    int num_runs = mca->nr;
+
+    // We always perform one more run than the num_runs
+    // value, so as to know "ne", i.e. the number of
+    // elements to reset in the T and F tables (labels and stats)
+    // After this first extra run, clock times are not accumulated
+    // and thus are lost.
+    // Note: the CLOCK_THREAD_START will still include this first run,
+    // and in case of multiple runs, only averaged times should be
+    // considered.
+    for (int run = 0; run < num_runs + 1; run++) {
+
+        CLOCK_THREAD_COMPUTE_START(mca->p);
+
+        MCA_Scatter_ImageX(mca);
+        pthread_barrier_wait(&main_barrier);
+
+        MCA_Label_Rosenfeld_PAR1(mca);
+        pthread_barrier_wait(&main_barrier);
+ 
 #if PARMERGE 
-    MCA_Label_Rosenfeld_PAR2(mca);
+        MCA_Label_Rosenfeld_PAR2(mca);
 #else
-    MCA_Label_Rosenfeld_PYR2(mca);
+        MCA_Label_Rosenfeld_PYR2(mca);
 #endif
-    pthread_barrier_wait(&main_barrier);
-    
-    MCA_Label_Rosenfeld_PAR3(mca);
-    pthread_barrier_wait(&main_barrier);
-
-    MCA_Gather_ImageL(mca);
-    pthread_barrier_wait(&main_barrier);
-
-    CLOCK_THREAD_COMPUTE_END(mca->p);
+        pthread_barrier_wait(&main_barrier);
+ 
+        MCA_Label_Rosenfeld_PAR3(mca);
+        pthread_barrier_wait(&main_barrier);
+
+        MCA_Gather_ImageL(mca);
+        pthread_barrier_wait(&main_barrier);
+        
+        CLOCK_THREAD_COMPUTE_END(mca->p);
+
+        if (run == 0) {
+            // Mise à jour du ne_prev par chaque thread
+            mca->ne_prev = mca->ne;
+            mca->ne = 0;
+        }
+        else {
+            // Accumulation du temps COMPUTE et de toutes les STEP
+            if (mca->p == 0) {
+                CLOCK_ACCUMULATE;
+            }
+            assert(mca->ne == mca->ne_prev);
+            // Reinitialisation de "ne" s'il ne s'agit pas du dernier run
+            if (run != num_runs) {
+                mca->ne = 0;
+            }
+        }
+        pthread_barrier_wait(&main_barrier);
+    }
+
  
 #if FEATURES