aws · aws-nslick · Aug 12, 2024 · Aug 12, 2024 · Aug 12, 2024 · Aug 13, 2024
diff --git a/include/nccl-headers/neuron/error.h b/include/nccl-headers/neuron/error.h
@@ -6,7 +6,7 @@
 #ifndef NCCL_HEADERS_NEURON_ERROR_H_
 #define NCCL_HEADERS_NEURON_ERROR_H_
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -21,7 +21,7 @@ typedef enum {
 	ncclRemoteError = 6
 } ncclResult_t;
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 }
 #endif
 

diff --git a/include/nccl-headers/neuron/net.h b/include/nccl-headers/neuron/net.h
@@ -6,7 +6,7 @@
 #ifndef NCCL_HEADERS_NEURON_NET_H_
 #define NCCL_HEADERS_NEURON_NET_H_
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -100,7 +100,7 @@ typedef struct {
 	ncclResult_t (*closeListen)(void* listenComm);
 } ncclNet_v4_t;
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 } // End extern "C"
 #endif
 

diff --git a/include/nccl_ofi.h b/include/nccl_ofi.h
@@ -5,7 +5,8 @@
 #ifndef NCCL_OFI_H_
 #define NCCL_OFI_H_
 
-#ifdef _cplusplus
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -23,7 +24,6 @@ extern "C" {
 #include "nccl_ofi_idpool.h"
 #include "nccl_ofi_mr.h"
 
-
 /*
  * NCCL_NET_HANDLE_MAXSIZE is a limited resource (and defined in NCCL).
  * An endpoint address buffer of 56 bytes *should* be large enough to hold
@@ -192,8 +192,7 @@ typedef struct nccl_ofi_connection_info {
 	nccl_net_ofi_req_t* req;
 } nccl_ofi_connection_info_t;
 /* Since this is a message on the wire, check that it has the expected size */
-_Static_assert(sizeof(nccl_ofi_connection_info_t) == 80,
-	       "Wrong size for SENDRECV connect message");
+static_assert(sizeof(nccl_ofi_connection_info_t) == 80, "Wrong size for SENDRECV connect message");
 
 typedef struct nccl_net_ofi_conn_handle {
 	char ep_name[MAX_EP_ADDR];
@@ -590,7 +589,7 @@ int nccl_net_ofi_dealloc_mr_buffer(void *ptr, size_t size);
 int nccl_net_ofi_query_provider_capabilities(const struct fi_info *selected_provider,
 					     unsigned int num_providers);
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 } // End extern "C"
 #endif
 

diff --git a/include/nccl_ofi_api.h b/include/nccl_ofi_api.h
@@ -6,13 +6,14 @@
 #ifndef NET_OFI_API_H_
 #define NET_OFI_API_H_
 
-#include "nccl-headers/net.h"
-#include "nccl-headers/error.h"
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 extern "C" {
 #endif
 
+#include "nccl-headers/net.h"
+#include "nccl-headers/error.h"
+
 struct nccl_ofi_properties;
 
 ncclResult_t nccl_net_ofi_init(ncclDebugLogger_t logFunction);
@@ -42,7 +43,7 @@ ncclResult_t nccl_net_ofi_closeSend(void *sendComm);
 ncclResult_t nccl_net_ofi_closeRecv(void *recvComm);
 ncclResult_t nccl_net_ofi_closeListen(void *listenComm);
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 }
 #endif // End extern "C"
 

diff --git a/include/nccl_ofi_config_bottom.h b/include/nccl_ofi_config_bottom.h
@@ -9,7 +9,15 @@
 #define OFI_LIKELY(x)   __builtin_expect((x), 1)
 #define OFI_UNLIKELY(x) __builtin_expect((x), 0)
 
+#ifdef __cplusplus
+#define NCCL_OFI_EXPORT_SYMBOL __attribute__((visibility("default"))) [[maybe_unused]]
+#else
 #define NCCL_OFI_EXPORT_SYMBOL __attribute__((visibility("default")))
+#endif
+
+#ifndef __cplusplus
+#define static_assert _Static_assert
+#endif
 
 /* Maximum length of directory path */
 #ifdef HAVE_LINUX_LIMITS_H

diff --git a/include/nccl_ofi_cuda.h b/include/nccl_ofi_cuda.h
@@ -6,7 +6,8 @@
 #ifndef NCCL_OFI_CUDA_H_
 #define NCCL_OFI_CUDA_H_
 
-#ifdef _cplusplus
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -41,7 +42,7 @@ extern CUresult (*nccl_net_ofi_cuFlushGPUDirectRDMAWrites)(CUflushGPUDirectRDMAW
 extern void *nccl_net_ofi_cuFlushGPUDirectRDMAWrites;
 #endif
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 } // End extern "C"
 #endif
 

diff --git a/include/nccl_ofi_deque.h b/include/nccl_ofi_deque.h
@@ -5,7 +5,8 @@
 #ifndef NCCL_OFI_DEQUE_H
 #define NCCL_OFI_DEQUE_H
 
-#ifdef _cplusplus
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -22,21 +23,20 @@ extern "C" {
  * this structure as a black box. Critically, the caller must ensure the
  * contents of this structure are not modified while it is in the deque.
  */
-struct nccl_ofi_deque_elem_t {
+struct nccl_ofi_deque_elem {
 	/* Pointer to previous element */
-	struct nccl_ofi_deque_elem_t *prev;
+	struct nccl_ofi_deque_elem *prev;
 	/* Pointer to next element */
-	struct nccl_ofi_deque_elem_t *next;
+	struct nccl_ofi_deque_elem *next;
 };
-typedef struct nccl_ofi_deque_elem_t nccl_ofi_deque_elem_t;
 
 /*
  * Deque (doubly-ended queue) structure
  *
  * Core deque structure.  This should be considered opaque to users
  * of the deque interface
  */
-struct nccl_ofi_deque_t {
+struct nccl_ofi_deque {
 	/* "Head" of queue.
 	 * The queue is circular. An empty queue has only this element. In an empty
 	 * queue, head.prev and head.next point to head.
@@ -45,18 +45,17 @@ struct nccl_ofi_deque_t {
 	 * front. insert_back and insert_front add elements to these respective
 	 * locations.
 	 */
-	nccl_ofi_deque_elem_t head;
+	struct nccl_ofi_deque_elem head;
 	/* Lock for deque operations */
 	pthread_mutex_t lock;
 };
-typedef struct nccl_ofi_deque_t nccl_ofi_deque_t;
 
 /*
  * Initialize deque structure.
  *
  * @return zero on success, non-zero on non-success.
  */
-int nccl_ofi_deque_init(nccl_ofi_deque_t **deque_p);
+extern int nccl_ofi_deque_init(struct nccl_ofi_deque **deque_p);
 
 /*
  * Finalize a deque
@@ -65,15 +64,15 @@ int nccl_ofi_deque_init(nccl_ofi_deque_t **deque_p);
  *
  * @return zero on success, non-zero on non-success.
  */
-int nccl_ofi_deque_finalize(nccl_ofi_deque_t *deque);
+extern int nccl_ofi_deque_finalize(struct nccl_ofi_deque *deque);
 
 /*
  * Insert an element to the back of the deque
  *
  * @param deque_elem	user-allocated storage space for list entry
  * @return zero on success, non-zero on error
  */
-static inline int nccl_ofi_deque_insert_back(nccl_ofi_deque_t *deque, nccl_ofi_deque_elem_t *deque_elem)
+static inline int nccl_ofi_deque_insert_back(struct nccl_ofi_deque *deque, struct nccl_ofi_deque_elem *deque_elem)
 {
 	assert(deque);
 	assert(deque_elem);
@@ -98,7 +97,7 @@ static inline int nccl_ofi_deque_insert_back(nccl_ofi_deque_t *deque, nccl_ofi_d
  * @param deque_elem	user-allocated storage space for list entry
  * @return zero on success, non-zero on error
  */
-static inline int nccl_ofi_deque_insert_front(nccl_ofi_deque_t *deque, nccl_ofi_deque_elem_t *deque_elem)
+static inline int nccl_ofi_deque_insert_front(struct nccl_ofi_deque *deque, struct nccl_ofi_deque_elem *deque_elem)
 {
 	assert(deque);
 	assert(deque_elem);
@@ -122,7 +121,7 @@ static inline int nccl_ofi_deque_insert_front(nccl_ofi_deque_t *deque, nccl_ofi_
  *
  * @return true if empty, false if not
  */
-static inline bool nccl_ofi_deque_isempty(nccl_ofi_deque_t *deque)
+static inline bool nccl_ofi_deque_isempty(struct nccl_ofi_deque *deque)
 {
 	return deque->head.next == &deque->head;
 }
@@ -132,7 +131,7 @@ static inline bool nccl_ofi_deque_isempty(nccl_ofi_deque_t *deque)
  * @param deque_elem  returned element; NULL if deque is empty or an error occurred
  * @return zero on success, non-zero on non-success
  */
-static inline int nccl_ofi_deque_remove_front(nccl_ofi_deque_t *deque, nccl_ofi_deque_elem_t **deque_elem)
+static inline int nccl_ofi_deque_remove_front(struct nccl_ofi_deque *deque, struct nccl_ofi_deque_elem **deque_elem)
 {
 	assert(deque);
 	assert(deque_elem);
@@ -162,7 +161,10 @@ static inline int nccl_ofi_deque_remove_front(nccl_ofi_deque_t *deque, nccl_ofi_
 	return 0;
 }
 
-#ifdef _cplusplus
+typedef struct nccl_ofi_deque_elem nccl_ofi_deque_elem_t;
+typedef struct nccl_ofi_deque nccl_ofi_deque_t;
+
+#ifdef __cplusplus
 } // End extern "C"
 #endif
 

diff --git a/include/nccl_ofi_ep_addr_list.h b/include/nccl_ofi_ep_addr_list.h
@@ -5,6 +5,10 @@
 #ifndef NCCL_OFI_EP_ADDR_LIST_H
 #define NCCL_OFI_EP_ADDR_LIST_H
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 struct nccl_ofi_ep_addr_list;
 typedef struct nccl_ofi_ep_addr_list nccl_ofi_ep_addr_list_t;
 
@@ -76,4 +80,8 @@ int nccl_ofi_ep_addr_list_delete(nccl_ofi_ep_addr_list_t *ep_list, nccl_net_ofi_
  */
 void nccl_ofi_ep_addr_list_fini(nccl_ofi_ep_addr_list_t *ep_list);
 
+#ifdef __cplusplus
+} // End extern "C"
+#endif
+
 #endif
diff --git a/include/nccl_ofi_freelist.h b/include/nccl_ofi_freelist.h
@@ -5,7 +5,8 @@
 #ifndef NCCL_OFI_FREELIST_H
 #define NCCL_OFI_FREELIST_H
 
-#ifdef _cplusplus
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -94,9 +95,9 @@ struct nccl_ofi_freelist_reginfo_t {
 };
 typedef struct nccl_ofi_freelist_reginfo_t nccl_ofi_freelist_reginfo_t;
 
-_Static_assert(offsetof(nccl_ofi_freelist_reginfo_t, elem) == 0,
+static_assert(offsetof(nccl_ofi_freelist_reginfo_t, elem) == 0,
 	       "elem is not the first member of the structure nccl_ofi_freelist_reginfo_t");
-_Static_assert(sizeof(nccl_ofi_freelist_reginfo_t) - offsetof(nccl_ofi_freelist_reginfo_t, redzone) == MEMCHECK_REDZONE_SIZE,
+static_assert(sizeof(nccl_ofi_freelist_reginfo_t) - offsetof(nccl_ofi_freelist_reginfo_t, redzone) == MEMCHECK_REDZONE_SIZE,
 	       "redzone is not the last member of the structure nccl_ofi_freelist_reginfo_t");
 
 /*
@@ -206,20 +207,20 @@ static inline void nccl_ofi_freelist_entry_set_undefined(nccl_ofi_freelist_t *fr
 		size_t redzone_offset = offsetof(struct nccl_ofi_freelist_reginfo_t, redzone);
 
 		/* Entry after reginfo structure is accessible but undefined */
-		nccl_net_ofi_mem_undefined_unaligned(entry_p + reginfo_offset + reginfo_size,
+		nccl_net_ofi_mem_undefined_unaligned((void*)((uintptr_t)entry_p + reginfo_offset + reginfo_size),
 						     user_entry_size - reginfo_offset - reginfo_size);
 		/* Redzone at the end of the reginfo structure is
 		 * marked as not accessible */
-		nccl_net_ofi_mem_noaccess_unaligned(entry_p + reginfo_offset + redzone_offset,
+		nccl_net_ofi_mem_noaccess_unaligned((void*)((uintptr_t)entry_p + reginfo_offset + redzone_offset),
 						    MEMCHECK_REDZONE_SIZE);
 		/* Members of reginfo structure except first and last
 		 * member are accessible and defined */
-		nccl_net_ofi_mem_defined_unaligned(entry_p + reginfo_offset + elem_size,
+		nccl_net_ofi_mem_defined_unaligned((void*)((uintptr_t)entry_p + reginfo_offset + elem_size),
 						   redzone_offset - elem_size);
 		/* First member of reginfo structure, i.e.,
 		 * nccl_ofi_freelist_elem_t structure, is marked as
 		 * not accessible */
-		nccl_net_ofi_mem_noaccess_unaligned(entry_p + reginfo_offset, elem_size);
+		nccl_net_ofi_mem_noaccess_unaligned((void*)((uintptr_t)entry_p + reginfo_offset), elem_size);
 		/* First part of entry until reginfo structure is
 		 * accessible but undefined */
 		nccl_net_ofi_mem_undefined(entry_p, reginfo_offset);
@@ -313,7 +314,7 @@ static inline void nccl_ofi_freelist_entry_free(nccl_ofi_freelist_t *freelist, v
 	nccl_net_ofi_mutex_unlock(&freelist->lock);
 }
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 } // End extern "C"
 #endif
 

diff --git a/include/nccl_ofi_idpool.h b/include/nccl_ofi_idpool.h
@@ -5,12 +5,12 @@
 #ifndef NCCL_OFI_IDPOOL_H_
 #define NCCL_OFI_IDPOOL_H_
 
-#ifdef _cplusplus
+
+#ifdef __cplusplus
 extern "C" {
 #endif
 
 #include <pthread.h>
-
 #include <stdint.h>
 
 /*
@@ -87,7 +87,7 @@ int nccl_ofi_idpool_free_id(nccl_ofi_idpool_t *idpool, int id);
  */
 int nccl_ofi_idpool_fini(nccl_ofi_idpool_t *idpool);
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 } // End extern "C"
 #endif
 

diff --git a/include/nccl_ofi_log.h b/include/nccl_ofi_log.h
@@ -6,7 +6,7 @@
 #ifndef NCCL_OFI_LOG_H_
 #define NCCL_OFI_LOG_H_
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -33,7 +33,7 @@ extern ncclDebugLogger_t ofi_log_function;
 #define NCCL_OFI_TRACE(flags, fmt, ...)
 #endif
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 } // End extern "C"
 #endif
 

diff --git a/include/nccl_ofi_math.h b/include/nccl_ofi_math.h
@@ -5,7 +5,7 @@
 #ifndef NCCL_OFI_MATH_H_
 #define NCCL_OFI_MATH_H_
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 extern "C" {
 #endif
 
@@ -65,7 +65,7 @@ extern "C" {
  */
 #define NCCL_OFI_ROUND_UP(x, y) NCCL_OFI_ROUND_DOWN((x) + ((y) - 1), (y))
 
-#ifdef _cplusplus
+#ifdef __cplusplus
 } // End extern "C"
 #endif