Update with nccl-2.23.4-1

include/core.h

@@ -50,44 +50,60 @@
 
 #include <errno.h>
 // Check system calls
-#define SYSCHECK(call, name) do { \
+#define SYSCHECK(statement, name) do { \
   int retval; \
-  SYSCHECKVAL(call, name, retval); \
-} while (false)
-
-#define SYSCHECKVAL(call, name, retval) do { \
-  SYSCHECKSYNC(call, name, retval); \
+  SYSCHECKSYNC((statement), name, retval); \
   if (retval == -1) { \
-    WARN("Call to " name " failed : %s", strerror(errno)); \
+    WARN("Call to " name " failed: %s", strerror(errno)); \
     return ncclSystemError; \
   } \
 } while (false)
 
-#define SYSCHECKSYNC(call, name, retval) do { \
-  retval = call; \
+#define SYSCHECKSYNC(statement, name, retval) do { \
+  retval = (statement); \
   if (retval == -1 && (errno == EINTR || errno == EWOULDBLOCK || errno == EAGAIN)) { \
     INFO(NCCL_ALL,"Call to " name " returned %s, retrying", strerror(errno)); \
   } else { \
     break; \
   } \
 } while(true)
 
-#define SYSCHECKGOTO(statement, RES, label) do { \
-  if ((statement) == -1) {    \
-    /* Print the back trace*/ \
-    RES = ncclSystemError;    \
-    INFO(NCCL_ALL,"%s:%d -> %d (%s)", __FILE__, __LINE__, RES, strerror(errno));    \
+#define SYSCHECKGOTO(statement, name, RES, label) do { \
+  int retval; \
+  SYSCHECKSYNC((statement), name, retval); \
+  if (retval == -1) { \
+    WARN("Call to " name " failed: %s", strerror(errno)); \
+    RES = ncclSystemError; \
     goto label; \
   } \
-} while (0);
+} while (0)
+
+// Pthread calls don't set errno and never return EINTR.
+#define PTHREADCHECK(statement, name) do { \
+  int retval = (statement); \
+  if (retval != 0) { \
+    WARN("Call to " name " failed: %s", strerror(retval)); \
+    return ncclSystemError; \
+  } \
+} while (0)
+
+#define PTHREADCHECKGOTO(statement, name, RES, label) do { \
+  int retval = (statement); \
+  if (retval != 0) { \
+    WARN("Call to " name " failed: %s", strerror(retval)); \
+    RES = ncclSystemError; \
+    goto label; \
+  } \
+} while (0)
+
 
 #define NEQCHECK(statement, value) do {   \
   if ((statement) != value) {             \
     /* Print the back trace*/             \
     INFO(NCCL_ALL,"%s:%d -> %d (%s)", __FILE__, __LINE__, ncclSystemError, strerror(errno));    \
     return ncclSystemError;     \
   }                             \
-} while (0);
+} while (0)
 
 #define NEQCHECKGOTO(statement, value, RES, label) do { \
   if ((statement) != value) { \
@@ -96,15 +112,15 @@
     INFO(NCCL_ALL,"%s:%d -> %d (%s)", __FILE__, __LINE__, RES, strerror(errno));    \
     goto label; \
   } \
-} while (0);
+} while (0)
 
 #define EQCHECK(statement, value) do {    \
   if ((statement) == value) {             \
     /* Print the back trace*/             \
     INFO(NCCL_ALL,"%s:%d -> %d (%s)", __FILE__, __LINE__, ncclSystemError, strerror(errno));    \
     return ncclSystemError;     \
   }                             \
-} while (0);
+} while (0)
 
 #define EQCHECKGOTO(statement, value, RES, label) do { \
   if ((statement) == value) { \
@@ -113,7 +129,7 @@
     INFO(NCCL_ALL,"%s:%d -> %d (%s)", __FILE__, __LINE__, RES, strerror(errno));    \
     goto label; \
   } \
-} while (0);
+} while (0)
 
 // Propagate errors up
 #define NCCLCHECK(call) do { \
@@ -122,15 +138,15 @@
     /* Print the back trace*/ \
     return RES; \
   } \
-} while (0);
+} while (0)
 
 #define NCCLCHECKGOTO(call, RES, label) do { \
   RES = call; \
   if (RES != ncclSuccess && RES != ncclInProgress) { \
     /* Print the back trace*/ \
     goto label; \
   } \
-} while (0);
+} while (0)
 
 #define NCCLWAIT(call, cond, abortFlagPtr) do {         \
   volatile uint32_t* tmpAbortFlag = (abortFlagPtr);     \
@@ -139,7 +155,7 @@
     return ncclInternalError;             \
   }                                       \
   if (tmpAbortFlag) NEQCHECK(*tmpAbortFlag, 0); \
-} while (!(cond));
+} while (!(cond))
 
 #define NCCLWAITGOTO(call, cond, abortFlagPtr, RES, label) do { \
   volatile uint32_t* tmpAbortFlag = (abortFlagPtr);             \
@@ -148,7 +164,7 @@
     goto label;                           \
   }                                       \
   if (tmpAbortFlag) NEQCHECKGOTO(*tmpAbortFlag, 0, RES, label); \
-} while (!(cond));
+} while (!(cond))
 
 #define NCCLCHECKTHREAD(a, args) do { \
   if (((args)->ret = (a)) != ncclSuccess && (args)->ret != ncclInProgress) { \

include/p2p_plugin.h

@@ -53,8 +53,13 @@ struct ncclIbMergedDev {
   int devs[NCCL_IB_MAX_DEVS_PER_NIC]; // Points to an index in ncclIbDevs
   int speed;
   char devName[MAX_MERGED_DEV_NAME]; // Up to NCCL_IB_MAX_DEVS_PER_NIC * name size, and a character for each '+'
+  int dmaBufSupported;               //  0 = uninit, 1 = yes, -1 = no
 } __attribute__((aligned(64)));
 
+struct ncclIbStats {
+  int fatalErrorCount;
+};
+
 struct ncclIbRequest {
   struct ncclIbNetCommBase* base;
   int type;
@@ -108,6 +113,7 @@ typedef struct ncclIbDev {
   struct   ncclIbMrCache mrCache;
   int ar; // ADAPTIVE_ROUTING
   struct ibv_port_attr portAttr;
+  struct ncclIbStats stats;
 } __attribute__((aligned(64))) ncclIbDev;
 
 
@@ -144,4 +150,6 @@ int ncclIbRelaxedOrderingCapable(void);
 
 nccl_p2p_plugin_t nccl_p2p_get_plugin_type();
 
+ncclResult_t ncclIbStatsInit(struct ncclIbStats* stat);
+
 #endif

include/timer.h

@@ -33,15 +33,15 @@ static double startTimes[8];
 #define TIME_START(index) do { \
   counts[index]++; \
   startTimes[index] = gettime(); \
-} while (0);
+} while (0)
 
 #define TIME_STOP(index) do { \
   times[index] += gettime() - startTimes[index]; \
-} while (0);
+} while (0)
 
 #define TIME_CANCEL(index) do { \
   counts[index]--; \
-} while (0);
+} while (0)
 
 #define TIME_PRINT(name) do { \
   printf("%s stats", name); \
@@ -50,11 +50,11 @@ static double startTimes[8];
     counts[i] = 0; \
   } \
   printf("\n"); \
-} while (0);
+} while (0)
 #else
-#define TIME_START(index) do {} while(0);
-#define TIME_STOP(index) do {} while(0);
-#define TIME_CANCEL(index) do {} while(0);
+#define TIME_START(index) do {} while(0)
+#define TIME_STOP(index) do {} while(0)
+#define TIME_CANCEL(index) do {} while(0)
 #define TIME_PRINT(name)
 #endif
 #endif