Merge release-7.2.0 (#171)

Merge release-7.2.0: * New default clock: get-time-of-day (#170) * STime API (#159) * Add ARMv7l arch support to ROSS (#155) * Fix to generate covage stats (#150) * changing damaris submodule/directory to risa * Updated README * Update to the way the build process grabs version number (#148) * simplifying the build of static or shared libraries (#147) * ROSS cleanup: -Wall, -Wextra, and more (#135)
ROSS-org · Dec 2, 2019 · 5eb95cf · 5eb95cf
2 parents adfaf05 + dc562c5
commit 5eb95cf
Show file tree

Hide file tree

Showing 33 changed files with 416 additions and 221 deletions.
diff --git a/.travis.yml b/.travis.yml
@@ -33,6 +33,12 @@ script:
  - make
  - CTEST_OUTPUT_ON_FAILURE=1 make test
  - cd ..
+ - mkdir build-gtod && cd build-gtod
+ - MPICH_CC=clag cmake -DCOVERALLS=ON -DCMAKE_BUILD_TYPE=Debug -DCMAKE_C_FLAGS="-Wall -Wextra" -DROSS_BUILD_MODELS=ON -DROSS_CLOCK_OVERRIDE=ON ..
+ - make
+ - CTEST_OUTPUT_ON_FAILURE=1 ctest -R SCHED
+ - make coveralls
+ - cd ..
  - mkdir build && cd build
  - MPICH_CC=clang cmake -DCOVERALLS=ON -DCMAKE_BUILD_TYPE=Debug -DCMAKE_C_FLAGS="-Wall -Wextra" -DROSS_BUILD_MODELS=ON ..
  - make

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -73,6 +73,7 @@ SET(GVT mpi_allreduce)
 
 # Architecture setting and management
 SET(VALID_ARCH NO)
+OPTION(ROSS_CLOCK_OVERRIDE "override platform detection to use gtod clock" NO)
 
 IF(${CMAKE_SYSTEM_PROCESSOR} STREQUAL i386)
   SET(VALID_ARCH YES)
@@ -132,16 +133,24 @@ IF(${CMAKE_SYSTEM_PROCESSOR} STREQUAL aarch64)
   SET(CLOCK aarch64)
 ENDIF(${CMAKE_SYSTEM_PROCESSOR} STREQUAL aarch64)
 
+IF(${CMAKE_SYSTEM_PROCESSOR} STREQUAL armv7l)
+  SET(VALID_ARCH YES)
+  SET(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+  ADD_DEFINITIONS(-D_GNU_SOURCE)
+  SET(CLOCK armv7l)
+ENDIF(${CMAKE_SYSTEM_PROCESSOR} STREQUAL armv7l)
 
-IF(VALID_ARCH)
+IF(VALID_ARCH AND NOT ${ROSS_CLOCK_OVERRIDE})
   MESSAGE(STATUS "System architecture detected: ${CMAKE_SYSTEM_PROCESSOR}")
   MESSAGE(STATUS "Using C_FLAGS: ${CMAKE_C_FLAGS}")
   MESSAGE(STATUS "Using CLOCK: ${CLOCK}")
-ELSE(VALID_ARCH)
-  MESSAGE(FATAL_ERROR "System architecture not recognized!\n"
+ELSE(VALID_ARCH AND NOT ${ROSS_CLOCK_OVERRIDE})
+  MESSAGE(WARNING "System architecture not recognized!\n"
     "Found: ${CMAKE_SYSTEM_PROCESSOR}\n"
-    "Expected: i386 | x86_64 | ppc64 | ppc64le | aarch64")
-ENDIF(VALID_ARCH)
+    "Falling back to get-time-of-day clock implementation.")
+  SET(CLOCK gtod)
+ENDIF(VALID_ARCH AND NOT ${ROSS_CLOCK_OVERRIDE})
+
 
 ## MPI
 INCLUDE(SetupMPI)

diff --git a/core/avl_tree.c b/core/avl_tree.c
@@ -42,8 +42,8 @@ avlSearch(AvlTree t, tw_event *key)
     if (t == AVL_EMPTY) {
         return 0;
     }
-    
-    if (key->recv_ts == t->key->recv_ts) {
+
+    if (TW_STIME_CMP(key->recv_ts, t->key->recv_ts) == 0) {
         // Timestamp is the same
         if (key->event_id == t->key->event_id) {
             // Event ID is the same
@@ -63,7 +63,7 @@ avlSearch(AvlTree t, tw_event *key)
     }
     else {
         // Timestamp is different
-        return avlSearch(t->child[key->recv_ts > t->key->recv_ts], key);
+        return avlSearch(t->child[TW_STIME_CMP(key->recv_ts, t->key->recv_ts) > 0], key);
     }
 }
 
@@ -72,12 +72,12 @@ void
 avlSanityCheck(AvlTree root)
 {
     int i;
-    
+
     if (root != AVL_EMPTY) {
         for (i = 0; i < 2; i++) {
             avlSanityCheck(root->child[i]);
         }
-        
+
         assert(root->height == 1 + ROSS_MAX(avlGetHeight(root->child[0]), avlGetHeight(root->child[1])));
     }
 }
@@ -87,7 +87,7 @@ static void
 avlFixHeight(AvlTree t)
 {
     assert(t != AVL_EMPTY);
-    
+
     t->height = 1 + ROSS_MAX(avlGetHeight(t->child[0]), avlGetHeight(t->child[1]));
 }
 
@@ -108,15 +108,15 @@ avlRotate(AvlTree *root, int d)
     AvlTree oldRoot;
     AvlTree newRoot;
     AvlTree oldMiddle;
-    
+
     oldRoot = *root;
     newRoot = oldRoot->child[d];
     oldMiddle = newRoot->child[!d];
-    
+
     oldRoot->child[d] = oldMiddle;
     newRoot->child[!d] = oldRoot;
     *root = newRoot;
-    
+
     /* update heights */
     avlFixHeight((*root)->child[!d]);   /* old root */
     avlFixHeight(*root);                /* new root */
@@ -129,7 +129,7 @@ static void
 avlRebalance(AvlTree *t)
 {
     int d;
-    
+
     if (*t != AVL_EMPTY) {
         for (d = 0; d < 2; d++) {
             /* maybe child[d] is now too tall */
@@ -146,11 +146,11 @@ avlRebalance(AvlTree *t)
                     avlRotate(&(*t)->child[d], !d);
                     avlRotate(t, d);
                 }
-                
+
                 return;   /* avlRotate called avlFixHeight */
             }
         }
-        
+
         /* update height */
         avlFixHeight(*t);
     }
@@ -171,19 +171,19 @@ avlInsert(AvlTree *t, tw_event *key)
                      "AVL tree out of memory.\nIncrease avl-size beyond %d\n",
                      (int)log2(g_tw_avl_node_count));
         }
-        
+
         (*t)->child[0] = AVL_EMPTY;
         (*t)->child[1] = AVL_EMPTY;
-        
+
         (*t)->key = key;
-        
+
         (*t)->height = 1;
-        
+
         /* done */
         return;
     }
-    
-    if (key->recv_ts == (*t)->key->recv_ts) {
+
+    if (TW_STIME_CMP(key->recv_ts, (*t)->key->recv_ts) == 0) {
         // We have a timestamp tie, check the event ID
         if (key->event_id == (*t)->key->event_id) {
             // We have a event ID tie, check the send_pe
@@ -203,7 +203,7 @@ avlInsert(AvlTree *t, tw_event *key)
     }
     else {
         // Timestamps are different
-        avlInsert(&(*t)->child[key->recv_ts > (*t)->key->recv_ts], key);
+        avlInsert(&(*t)->child[TW_STIME_CMP(key->recv_ts, (*t)->key->recv_ts) > 0], key);
         avlRebalance(t);
     }
 }
@@ -227,9 +227,9 @@ avlDeleteMin(AvlTree *t)
 {
     AvlTree oldroot;
     tw_event *event_with_lowest_ts = NULL;
-    
+
     assert(t != AVL_EMPTY);
-    
+
     if ((*t)->child[0] == AVL_EMPTY) {
         /* root is min value */
         oldroot = *t;
@@ -241,7 +241,7 @@ avlDeleteMin(AvlTree *t)
         /* min value is in left subtree */
         event_with_lowest_ts = avlDeleteMin(&(*t)->child[0]);
     }
-    
+
     avlRebalance(t);
     return event_with_lowest_ts;
 }
@@ -252,13 +252,13 @@ avlDelete(AvlTree *t, tw_event *key)
 {
     tw_event *target = NULL;
     AvlTree oldroot;
-    
+
     if (*t == AVL_EMPTY) {
         tw_error(TW_LOC, "We never look for non-existent events!");
         return target;
     }
-    
-    if (key->recv_ts == (*t)->key->recv_ts) {
+
+    if (TW_STIME_CMP(key->recv_ts, (*t)->key->recv_ts) == 0) {
         // We have a timestamp tie, check the event ID
         if (key->event_id == (*t)->key->event_id) {
             // We have a event ID tie, check the send_pe
@@ -289,27 +289,27 @@ avlDelete(AvlTree *t, tw_event *key)
     }
     else {
         // Timestamps are different
-        target = avlDelete(&(*t)->child[key->recv_ts > (*t)->key->recv_ts], key);
+        target = avlDelete(&(*t)->child[TW_STIME_CMP(key->recv_ts, (*t)->key->recv_ts) > 0], key);
     }
-    
+
     avlRebalance(t);
-    
+
     return target;
 }
 
 AvlTree avl_alloc(void)
 {
     AvlTree head = g_tw_pe->avl_list_head;
     g_tw_pe->avl_list_head = head->next;
-    
+
     if (g_tw_pe->avl_list_head == NULL) {
         tw_error(TW_LOC,
                  "AVL tree out of memory.\nIncrease avl-size beyond %d\n",
                  (int)log2(g_tw_avl_node_count));
     }
-    
+
     head->next = NULL;
-    
+
     return head;
 }
 

diff --git a/core/clock/armv7l.c b/core/clock/armv7l.c
@@ -0,0 +1,48 @@
+/*
+	This implementation of an ARM v7 clock reader utilizes the
+	Performance Monitoring Unit (PMU) on Cortex-A7 chips.
+	Unfortunately, access to the cycle counter from userspace
+	is disabled by default. A kernel module that enables access
+	from userspace is required or the system will fault.
+
+	An example kernel module that does just that can be found:
+	https://github.com/nmcglohon/armv7l-userspace-counter.git
+
+	More information can be found:
+	http://neocontra.blogspot.com/2013/05/user-mode-performance-counters-for.html
+ */
+
+#include <ross.h>
+
+#ifndef __GNUC__
+#  error gcc asm extensions required
+#endif
+#if ! (defined(__arm__))
+#  error only 32 bit arm platform supported
+#endif
+
+static const tw_optdef clock_opts [] =
+{
+	TWOPT_GROUP("ROSS Timing"),
+	TWOPT_STIME("clock-rate", g_tw_clock_rate, "CPU Clock Rate"),
+	TWOPT_END()
+};
+
+const tw_optdef *tw_clock_setup(void)
+{
+	return clock_opts;
+}
+
+
+void tw_clock_init(tw_pe * me)
+{
+	me->clock_time = 0;
+	me->clock_offset = tw_clock_read();
+}
+
+
+tw_clock tw_clock_now(tw_pe * me)
+{
+	me->clock_time = tw_clock_read() - me->clock_offset;
+	return me->clock_time;
+}
diff --git a/core/clock/armv7l.h b/core/clock/armv7l.h
@@ -0,0 +1,32 @@
+/*
+	This implementation of an ARM v7 clock reader utilizes the
+	Performance Monitoring Unit (PMU) on Cortex-A7 chips.
+	Unfortunately, access to the cycle counter from userspace
+	is disabled by default. A kernel module that enables access
+	from userspace is required or the system will fault.
+
+	An example kernel module that does just that can be found:
+	https://github.com/nmcglohon/armv7l-userspace-counter.git
+
+	More information can be found:
+	http://neocontra.blogspot.com/2013/05/user-mode-performance-counters-for.html
+ */
+
+#ifndef INC_clock_armv7l
+#define INC_clock_armv7l
+
+typedef unsigned int tw_clock;
+
+static inline tw_clock  tw_clock_read(void)
+{
+	unsigned int result;
+#ifdef ROSS_timing
+	do {
+		__asm__ __volatile__ ("MRC p15, 0, %0, c9, c13, 0" : "=r"(result));
+	} while (__builtin_expect ((int) result == -1, 0));
+#endif
+
+	return result;
+}
+
+#endif
diff --git a/core/clock/gtod.c b/core/clock/gtod.c
@@ -0,0 +1,53 @@
+#include <ross.h>
+
+extern unsigned long long g_tw_clock_rate;
+
+static const tw_optdef clock_opts [] =
+{
+ TWOPT_GROUP("ROSS Timing"),
+ TWOPT_ULONGLONG("clock-rate", g_tw_clock_rate, "CPU Clock Rate"),
+ TWOPT_END()
+};
+
+const tw_optdef *tw_clock_setup(void)
+{
+	return clock_opts;
+}
+
+tw_clock tw_clock_read(void)
+{
+#ifdef ZERO_BASED
+  static volatile int inited = 0;
+  static volatile tw_clock base = 0;
+#else
+  const tw_clock base = 0;
+#endif
+
+  const tw_clock scale = 1000000;
+  struct timeval tv;
+  gettimeofday(&tv,NULL);
+
+#ifdef ZERO_BASED
+  if(inited == 0) {
+    base = ((tw_clock) tv.tv_sec)*scale + (tw_clock) tv.tv_usec;
+    inited = 1;
+  }
+#endif
+
+  return
+    (((tw_clock) tv.tv_sec)*scale + (tw_clock) tv.tv_usec) - base;
+}
+
+void
+tw_clock_init(tw_pe * me)
+{
+	me->clock_time = 0;
+	me->clock_offset = tw_clock_read();
+}
+
+tw_clock
+tw_clock_now(tw_pe * me)
+{
+	me->clock_time = tw_clock_read() - me->clock_offset;
+	return me->clock_time;
+}
diff --git a/core/clock/gtod.h b/core/clock/gtod.h
@@ -0,0 +1,6 @@
+#ifndef INC_clock_gtod
+#define INC_clock_gtod
+
+typedef uint64_t tw_clock;
+
+#endif