Changeset 630 for trunk/user/fft/fft.c

Timestamp:

May 21, 2019, 6:00:06 PM (5 years ago)

Author:

alain

Message:

1) Fix a bug in the vfs_add_special_dentries() function:
The <.> and <..> dentries must not be created on IOC and on the mapper
for the VFS root directory.
2) Fix a bug in the hal_gpt_allocate_pt2 function, related to the
use of the TSAR_LOCKED attribute to avoid concurrent mapping of the PTD1.

File:

• trunk/user/fft/fft.c (modified) (11 diffs)

trunk/user/fft/fft.c

@@ -91 +91 @@
 #define MODE                    COSIN           // DATA array initialisation mode
 #define CHECK                   0               
-#define DEBUG_MAIN              1               // trace main() function (detailed if odd)
+#define DEBUG_MAIN              0               // trace main() function (detailed if odd)
 #define DEBUG_SLAVE             0               // trace slave() function (detailed if odd)
 #define DEBUG_FFT1D             0               // trace FFT1D() function (detailed if odd)
@@ -404 +404 @@
     }
 
-
-    // register sequencial initalisation completion cycle
+    printf("\n[fft] main completes threads creation\n");
+
     get_cycle( &end_init_cycle );
+
+    // register sequencial time
     init_time = (unsigned int)(end_init_cycle - start_init_cycle);
-
-    printf("\n[fft] main completes threads creation\n");
     
     // main itself executes the slave() function
@@ -510 +510 @@
     unsigned int max_sync = sync_time[0];
 
-    for (tid = 1 ; tid < nthreads ; tid++) 
+    for (tid = 0 ; tid < nthreads ; tid++) 
     {
         if (parallel_time[tid] > max_para)  max_para = parallel_time[tid];
@@ -566 +566 @@
     MainNum = args->main_tid;
 
-    // initialise instrumentation
     get_cycle( &parallel_start );
 
@@ -578 +577 @@
     pthread_barrier_wait( &barrier );
     get_cycle( &barrier_stop );
-    sync_time[MyNum] += (barrier_stop - barrier_start);
-
-// printf("\n[@@@] %s : thread %d exit first barrier / cycle %d\n",
-// __FUNCTION__, MyNum, (unsigned int)barrier_stop );
+    sync_time[MyNum] += (unsigned int)(barrier_stop - barrier_start);
+
+#if DEBUG_SLAVE
+printf("\n[@@@] %s : thread %d exit first barrier / cycle %d\n",
+__FUNCTION__, MyNum, (unsigned int)barrier_stop );
+#endif
 
     // allocate and initialise local array upriv[] 
@@ -605 +606 @@
 pthread_barrier_wait( &barrier );
 get_cycle( &barrier_stop );
-sync_time[MyNum] += (long)(barrier_stop - barrier_start);
+sync_time[MyNum] += (unsigned int)(barrier_stop - barrier_start);
 FFT1D( -1 , data , trans , upriv , twid , MyNum , MyFirst , MyLast );
 #endif
 
-    // register computation time
     get_cycle( &parallel_stop );
-    parallel_time[MyNum] = (parallel_stop - parallel_start);
+
+    // register parallel time
+    parallel_time[MyNum] = (unsigned int)(parallel_stop - parallel_start);
 
 #if DEBUG_SLAVE
-printf("\n[fft] %s : thread %x exit / cycle %d\n", __FUNCTION__, MyNum, parallel_stop );
+printf("\n[fft] %s : thread %x completes fft / p_start %d / p_stop %d\n", 
+__FUNCTION__, MyNum, (unsigned int)parallel_start, (unsigned int)parallel_stop );
+int tid;
+for (tid = 0 ; tid < nthreads ; tid++) 
+{
+    printf("- tid %d : Sequencial %d / Parallel %d / Barrier %d\n",
+    tid , init_time, parallel_time[tid], sync_time[tid] );
+}
 #endif
 
@@ -867 +876 @@
     pthread_barrier_wait( &barrier );
     get_cycle( &barrier_stop );
-    sync_time[MyNum] = (long)(barrier_stop - barrier_start);
+    sync_time[MyNum] = (unsigned int)(barrier_stop - barrier_start);
 
 #if( DEBUG_FFT1D & 1 )
@@ -897 +906 @@
 #endif
 
-    sync_time[MyNum] += (long)(barrier_stop - barrier_start);
+    sync_time[MyNum] += (unsigned int)(barrier_stop - barrier_start);
 
     // transpose tmp to x
@@ -916 +925 @@
 #endif
 
-    sync_time[MyNum] += (long)(barrier_stop - barrier_start);
+    sync_time[MyNum] += (unsigned int)(barrier_stop - barrier_start);
 
     // do FFTs on rows of x and apply the scaling factor 
@@ -938 +947 @@
 printf("\n[fft] %s : thread %d exit barrier after FFT on rows\n", __FUNCTION__, MyNum);
 #endif
-    sync_time[MyNum] += (long)(barrier_stop - barrier_start);
+    sync_time[MyNum] += (unsigned int)(barrier_stop - barrier_start);
 
     // transpose x to tmp
@@ -957 +966 @@
 #endif
 
-    sync_time[MyNum] += (long)(barrier_stop - barrier_start);
+    sync_time[MyNum] += (unsigned int)(barrier_stop - barrier_start);
     sync_time[MyNum] += (long)(barrier_stop - barrier_start);