use C++ for unit tests

.github/workflows/distcheck.yaml

@@ -85,7 +85,7 @@ jobs:
 
       - name: Install hwloc, utilities.
         run: |
-          yum -y install hwloc-devel autoconf automake libtool gcc g++ git make
+          yum -y install hwloc-devel autoconf automake libtool gcc gcc-c++ git make
 
       - name: Install CUDA
         if: matrix.sdk == 'cuda'

configure.ac

@@ -46,6 +46,7 @@ NCCL_NET_OFI_DISTCHCK_CONFIGURE_FLAGS=
 # Checks for programs
 AC_PROG_CC
 AM_PROG_CC_C_O
+AC_PROG_CXX
 m4_version_prereq([2.70], [], [AC_PROG_CC_STDC])
 AC_C_INLINE
 
@@ -55,12 +56,18 @@ dnl between this and AC_PROG_CC.
 AS_IF([test "${ax_enable_debug}" = "no"], [
              dnl Enable O3 optimization
              CFLAGS="${CFLAGS} -O3"
+             CXXFLAGS="${CXXFLAGS} -O3"
+
+             dnl use C++17, disable exceptions, and disable runtime type information
+             CXXFLAGS="-std=c++17 -fno-rtti -fno-exceptions"
 
              dnl dead code elim
              CFLAGS="${CFLAGS} -ffunction-sections -fdata-sections"
+             CXXFLAGS="${CXXFLAGS} -ffunction-sections -fdata-sections"
 
              dnl https://maskray.me/blog/2021-05-09-fno-semantic-interposition
              CFLAGS="${CFLAGS} -fno-semantic-interposition -fvisibility=hidden"
+             CXXFLAGS="${CXXFLAGS} -fno-semantic-interposition -fvisibility=hidden "
              LDFLAGS="${LDFLAGS} -Bsymbolic"
             ])
 CHECK_ENABLE_SANITIZER()
@@ -175,10 +182,15 @@ AS_IF([test "${enable_trace}" = "yes" ],
 AC_DEFINE_UNQUOTED([OFI_NCCL_TRACE], [${trace}], [Defined to 1 unit test output should include TRACE level])
 
 picky_cflags=""
+picky_cxxflags=""
 AC_DEFUN([ADD_PICKY_FLAGS],[
     AC_LANG_PUSH([C])
     AX_CHECK_COMPILE_FLAG([$1], [picky_cflags="${picky_cflags} $1"], [], [-Werror])
     AC_LANG_POP()
+
+    AC_LANG_PUSH([C++])
+    AX_CHECK_COMPILE_FLAG([$1], [picky_cxxflags="${picky_cxxflags} $1"], [], [-Werror -x c++])
+    AC_LANG_POP()
     ])
 
 dnl standard and normal
@@ -220,17 +232,21 @@ AC_ARG_ENABLE([picky-compiler],
 AS_IF([test "${enable_picky_compiler}" != "no"],
       [AC_MSG_NOTICE([Adding ${picky_cflags} to CFLAGS.])
        CFLAGS="${CFLAGS} ${picky_cflags}"
+       CXXFLAGS="${CXXFLAGS} ${picky_cxxflags}"
+       AS_UNSET([picky_cxxflags])
        AS_UNSET([picky_cflags])])
 
 werror_flags="-Werror"
 AC_ARG_ENABLE([werror],
    [AS_HELP_STRING([--enable-werror], [(Developer) Enable setting -Werror.  Off by default, unless building from Git tree.])])
 AS_IF([test -d "${srcdir}/.git" -a -z "${enable_werror}"],
       [AC_MSG_NOTICE([Found .git directory.  Adding ${werror_flags} to CFLAGS.])
-       CFLAGS="${werror_flags} ${CFLAGS}"],
+       CXXFLAGS="${werror_flags} ${CXXFLAGS} "
+       CFLAGS="${werror_flags} ${CFLAGS} "],
       [test "${enable_werror}" = "yes"],
       [AC_MSG_NOTICE([Adding ${werror_flags} to CFLAGS.])
-       CFLAGS="${werror_flags} ${CFLAGS}"])
+       CXXFLAGS="${werror_flags} ${CXXFLAGS} "
+       CFLAGS="${werror_flags} ${CFLAGS} "])
 
 AC_SUBST([NCCL_NET_OFI_DISTCHCK_CONFIGURE_FLAGS])
 

include/nccl_ofi_config_bottom.h

@@ -5,6 +5,8 @@
 #ifndef NCCL_OFI_CONFIG_BOTTOM_H
 #define NCCL_OFI_CONFIG_BOTTOM_H
 
+#define NCCL_OFI_N_NVTX_DOMAIN_PER_COMM 8
+
 /* configure aborts if __buildin_expect() isn't available */
 #define OFI_LIKELY(x)   __builtin_expect((x), 1)
 #define OFI_UNLIKELY(x) __builtin_expect((x), 0)

include/nccl_ofi_mr.h

@@ -33,14 +33,15 @@ enum nccl_ofi_mr_ckey_type {
 typedef enum nccl_ofi_mr_ckey_type nccl_ofi_mr_ckey_type_t;
 
 struct nccl_ofi_mr_ckey {
-	const union {
-		const struct iovec iovec;
+	union {
+		struct iovec iovec;
 #if HAVE_DECL_FI_MR_DMABUF
-		const struct fi_mr_dmabuf fi_mr_dmabuf;
+		struct fi_mr_dmabuf fi_mr_dmabuf;
 #endif
 	};
-	const enum nccl_ofi_mr_ckey_type type;
+	enum nccl_ofi_mr_ckey_type type;
 };
+
 typedef struct nccl_ofi_mr_ckey nccl_ofi_mr_ckey_t;
 typedef struct nccl_ofi_mr_ckey const *const nccl_ofi_mr_ckey_ref;
 
@@ -118,30 +119,24 @@ static inline void nccl_ofi_mr_ckey_round(size_t *len, void **base_addr, const c
 #if HAVE_DECL_FI_MR_DMABUF
 static inline nccl_ofi_mr_ckey_t nccl_ofi_mr_ckey_mk_dmabuf(int fd, uint64_t offset, size_t len, void *base_addr)
 {
-	return (nccl_ofi_mr_ckey_t){
-		.fi_mr_dmabuf =
-			{
-				.fd = fd,
-				.offset = offset,
-				.len = len,
-				.base_addr = base_addr,
-			},
-		.type = NCCL_OFI_MR_CKEY_DMABUF,
-	};
+	nccl_ofi_mr_ckey_t cache_key = {};
+	cache_key.fi_mr_dmabuf.fd = fd;
+	cache_key.fi_mr_dmabuf.offset = offset;
+	cache_key.fi_mr_dmabuf.len = len;
+	cache_key.fi_mr_dmabuf.base_addr = base_addr;
+	cache_key.type = NCCL_OFI_MR_CKEY_DMABUF;
+	return cache_key;
 }
 #endif
 
 static inline nccl_ofi_mr_ckey_t nccl_ofi_mr_ckey_mk_vec(void *iov_base, size_t iov_len)
 {
 	nccl_ofi_mr_ckey_round(&iov_len, &iov_base, "iovec");
-	return (nccl_ofi_mr_ckey_t){
-		.iovec =
-			{
-				.iov_base = iov_base,
-				.iov_len = iov_len,
-			},
-		.type = NCCL_OFI_MR_CKEY_IOVEC,
-	};
+	nccl_ofi_mr_ckey_t cache_key = {};
+	cache_key.iovec.iov_base = iov_base;
+	cache_key.iovec.iov_len = iov_len;
+	cache_key.type = NCCL_OFI_MR_CKEY_IOVEC;
+	return cache_key;
 }
 
 static inline void nccl_ofi_mr_ckey_fill_mr_attrs(nccl_ofi_mr_ckey_ref ckey, struct fi_mr_attr *attrs, uint64_t *flags)

include/nccl_ofi_rdma.h

@@ -9,18 +9,22 @@
 extern "C" {
 #endif
 
+#include "config.h"
+
 #include <rdma/fabric.h>
 
 #include "nccl_ofi.h"
-#include "nccl_ofi_log.h"
-#include "nccl_ofi_scheduler.h"
-#include "nccl_ofi_msgbuff.h"
-#include "nccl_ofi_topo.h"
 #include "nccl_ofi_deque.h"
+#include "nccl_ofi_ep_addr_list.h"
 #include "nccl_ofi_freelist.h"
 #include "nccl_ofi_idpool.h"
-#include "nccl_ofi_tracepoint.h"
-#include "nccl_ofi_ep_addr_list.h"
+#include "nccl_ofi_log.h"
+#include "nccl_ofi_msgbuff.h"
+#include "nccl_ofi_scheduler.h"
+#include "nccl_ofi_topo.h"
+#if HAVE_NVTX_TRACING
+#include <nvtx3/nvToolsExt.h>
+#endif
 
 /* Maximum number of rails supported. This defines the size of
  * messages exchanged during connection establishment (linear

include/tracing_impl/nvtx.h

@@ -8,8 +8,6 @@
 #if HAVE_NVTX_TRACING
 #include <nvtx3/nvToolsExt.h>
 
-#define NCCL_OFI_N_NVTX_DOMAIN_PER_COMM 8
-
 static inline void nvtx_mark_domain(nvtxDomainHandle_t domain, const char* name, uint32_t color)
 {
 	nvtxEventAttributes_t eventAttrib = {};

m4/check_pkg_mpi.m4

@@ -31,13 +31,15 @@ AC_DEFUN([CHECK_PKG_MPI], [
          LDFLAGS="${MPI_LDFLAGS} ${LDFLAGS}"])
 
   MPICC=${mpi_bindir}mpicc
+  MPICXX=${mpi_bindir}mpicxx
 
   AC_MSG_CHECKING([for working mpicc])
   ${MPICC} --help >& AS_MESSAGE_LOG_FD
   AS_IF([test $? -eq 0],
         [AC_MSG_RESULT([yes])],
         [AC_MSG_RESULT([no])
          MPICC="${CC}"
+         MPICXX="${CXX}"
          AS_IF([test "${check_pkg_found}" = "yes"],
                [AC_CHECK_HEADERS([mpi.h], [], [check_pkg_found=no])])
 
@@ -49,6 +51,7 @@ AC_DEFUN([CHECK_PKG_MPI], [
         [$2])
 
   AC_SUBST([MPICC])
+  AC_SUBST([MPICXX])
   AC_SUBST([MPI_CPPFLAGS])
   AC_SUBST([MPI_LDFLAGS])
   AC_SUBST([MPI_LIBS])

tests/functional/Makefile.am

@@ -12,13 +12,14 @@ AM_CPPFLAGS += $(MPI_CPPFLAGS) $(CUDA_CPPFLAGS)
 AM_LDFLAGS = $(MPI_LDFLAGS) $(CUDA_LDFLAGS)
 LDADD = $(top_builddir)/src/libinternal_net_plugin.la $(MPI_LIBS) $(CUDA_LIBS)
 CC = $(MPICC)
+CXX = $(MPICXX)
 
 if ENABLE_FUNC_TESTS
-noinst_HEADERS = test-common.h
+noinst_HEADERS = test-common.hpp
 
 bin_PROGRAMS = nccl_connection nccl_message_transfer ring
 
-nccl_connection_SOURCES = nccl_connection.c
-nccl_message_transfer_SOURCES = nccl_message_transfer.c
-ring_SOURCES = ring.c
+nccl_connection_SOURCES = nccl_connection.cc
+nccl_message_transfer_SOURCES = nccl_message_transfer.cc
+ring_SOURCES = ring.cc
 endif

tests/functional/nccl_connection.c → tests/functional/nccl_connection.cc

@@ -8,7 +8,7 @@
 
 #include "config.h"
 
-#include "test-common.h"
+#include "test-common.hpp"
 
 int main(int argc, char* argv[])
 {

tests/functional/nccl_message_transfer.c → tests/functional/nccl_message_transfer.cc

@@ -9,14 +9,16 @@
 
 #include "config.h"
 
-#include "test-common.h"
+#include "test-common.hpp"
 
 #define PROC_NAME_IDX(i) (i * MPI_MAX_PROCESSOR_NAME)
 
 int main(int argc, char* argv[])
 {
 	ncclResult_t res = ncclSuccess;
-	int rank, proc_name_len, num_ranks = 0, local_rank = 0, peer_rank = 0;
+	int rank, proc_name_len, num_ranks = 0, peer_rank = 0;
+	/* Unused when trace prints are not enabled. */
+	[[maybe_unused]] int local_rank = 0;
 	int buffer_type = NCCL_PTR_HOST;
 	test_nccl_properties_t props = {};
 
@@ -390,7 +392,7 @@ int main(int argc, char* argv[])
 	MPI_Finalize();
 	NCCL_OFI_INFO(NCCL_NET, "Test completed successfully for rank %d", rank);
 
-exit:;
+exit:
 
 	ncclResult_t close_res = ncclSuccess;
 

tests/functional/ring.c → tests/functional/ring.cc

@@ -4,7 +4,7 @@
 
 #include "config.h"
 
-#include "test-common.h"
+#include "test-common.hpp"
 
 #define PROC_NAME_IDX(i) (i * MPI_MAX_PROCESSOR_NAME)
 

tests/functional/test-common.h → tests/functional/test-common.hpp

@@ -23,6 +23,10 @@
 #include "nccl_ofi_math.h"
 #include "nccl_ofi_param.h"
 
+
+template <typename T>
+constexpr T PROC_NAME_IDX(T i) { return i * MPI_MAX_PROCESSOR_NAME; };
+
 #define STR2(v)		#v
 #define STR(v)		STR2(v)
 

tests/unit/Makefile.am

@@ -9,7 +9,7 @@ AM_CPPFLAGS = -I$(top_srcdir)/include
 AM_CPPFLAGS += -isystem $(abs_top_srcdir)/3rd-party/nccl/$(DEVICE_INTERFACE)/include
 AM_CPPFLAGS += -isystem $(abs_top_srcdir)/3rd-party/uthash/include
 LDADD = $(top_builddir)/src/libinternal_net_plugin.la
-noinst_HEADERS = test-common.h
+noinst_HEADERS = test-common.hpp
 
 noinst_PROGRAMS = \
 	deque \
@@ -26,21 +26,21 @@ if WANT_PLATFORM_AWS
   AM_CPPFLAGS += $(CUDA_CPPFLAGS)
   LDADD += $(CUDA_LIBS)
   noinst_PROGRAMS += show_tuner_decisions
-  show_tuner_decisions_SOURCES = show_tuner_decisions.c
+  show_tuner_decisions_SOURCES = show_tuner_decisions.cc
   show_tuner_decisions_LDADD = $(top_builddir)/src/libinternal_tuner_plugin.la
   noinst_PROGRAMS += region_based_tuner
   region_based_tuner_SOURCES = region_based_tuner.c
   region_based_tuner_LDADD = $(top_builddir)/src/libinternal_tuner_plugin.la
 endif
 endif
 
-idpool_SOURCES = idpool.c
-deque_SOURCES = deque.c
-freelist_SOURCES = freelist.c
-msgbuff_SOURCES = msgbuff.c
-scheduler_SOURCES = scheduler.c
-ep_addr_list_SOURCES = ep_addr_list.c
-mr_SOURCES = mr.c
+idpool_SOURCES = idpool.cc
+deque_SOURCES = deque.cc
+freelist_SOURCES = freelist.cc
+msgbuff_SOURCES = msgbuff.cc
+scheduler_SOURCES = scheduler.cc
+ep_addr_list_SOURCES = ep_addr_list.cc
+mr_SOURCES = mr.cc
 
 TESTS = $(noinst_PROGRAMS)
 endif

tests/unit/deque.c → tests/unit/deque.cc

@@ -6,7 +6,7 @@
 
 #include <stdio.h>
 
-#include "test-common.h"
+#include "test-common.hpp"
 #include "nccl_ofi_deque.h"
 
 #define test_get_front(deque, expected) \

tests/unit/freelist.c → tests/unit/freelist.cc

@@ -6,7 +6,7 @@
 
 #include <stdio.h>
 
-#include "test-common.h"
+#include "test-common.hpp"
 #include "nccl_ofi_freelist.h"
 
 void *simple_base;

tests/unit/idpool.c → tests/unit/idpool.cc

@@ -6,7 +6,7 @@
 
 #include <stdio.h>
 
-#include "test-common.h"
+#include "test-common.hpp"
 #include "nccl_ofi_idpool.h"
 #include "nccl_ofi_math.h"
 

tests/unit/mr.c → tests/unit/mr.cc

@@ -6,7 +6,7 @@
 
 #include <stdlib.h>
 
-#include "test-common.h"
+#include "test-common.hpp"
 #include "nccl_ofi_mr.h"
 
 static inline bool test_lookup_impl(nccl_ofi_mr_cache_t *cache, void *addr, size_t size,

tests/unit/msgbuff.c → tests/unit/msgbuff.cc

@@ -8,7 +8,7 @@
 
 #include "nccl_ofi_msgbuff.h"
 
-#include "test-common.h"
+#include "test-common.hpp"
 
 int main(int argc, char *argv[])
 {