Fix some libraries caused conflicts (libdispatch v.s. nvml).

liuliu · Jun 7, 2024 · e63fd6e · e63fd6e
1 parent 78fedee
commit e63fd6e
Show file tree

Hide file tree

Showing 4 changed files with 97 additions and 95 deletions.
diff --git a/lib/configure b/lib/configure
@@ -3884,7 +3884,7 @@ printf "%s\n" "$ax_cv_check_cflags_fftw3_h" >&6; }
 	if test "x$ax_cv_check_cflags_fftw3_h" = xyes
 then :
   DEFINE_MACROS="$DEFINE_MACROS-D HAVE_FFTW3 "
- MKLDFLAGS="$MKLDFLAGS-lfftw3 -lfftw3f -lpthread "
+ MKLDFLAGS="$MKLDFLAGS-lfftw3 -lfftw3f "
 
 else $as_nop
   :
@@ -4300,90 +4300,6 @@ if test -n "${OPENMP_CFLAGS}"; then
 
 	MKLDFLAGS="$MKLDFLAGS$OPENMP_CFLAGS "
 
-fi
-# only check dispatch if compiled with clang
-if test $CC = clang; then
-	 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking dispatch/dispatch.h presence" >&5
-printf %s "checking dispatch/dispatch.h presence... " >&6; }
-if test ${ax_cv_check_cflags_dispatch_dispatch_h+y}
-then :
-  printf %s "(cached) " >&6
-else $as_nop
-
-		ax_check_save_flags=$CFLAGS
-		CFLAGS="$CFLAGS dispatch/dispatch.h"
-		cat confdefs.h - <<_ACEOF >conftest.$ac_ext
-/* end confdefs.h.  */
-#include <dispatch/dispatch.h>
-int
-main (void)
-{
-
-  ;
-  return 0;
-}
-_ACEOF
-if ac_fn_c_try_cpp "$LINENO"
-then :
-  ax_cv_check_cflags_dispatch_dispatch_h=yes
-else $as_nop
-  ax_cv_check_cflags_dispatch_dispatch_h=no
-fi
-rm -f conftest.err conftest.i conftest.$ac_ext
-		CFLAGS=$ax_check_save_flags
-fi
-{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_dispatch_dispatch_h" >&5
-printf "%s\n" "$ax_cv_check_cflags_dispatch_dispatch_h" >&6; }
-	if test "x$ax_cv_check_cflags_dispatch_dispatch_h" = xyes
-then :
-  DEFINE_MACROS="$DEFINE_MACROS-D USE_DISPATCH "
- MKCFLAGS="$MKCFLAGS-fblocks "
-
-else $as_nop
-  :
-fi
-
-	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for dispatch_apply in -ldispatch" >&5
-printf %s "checking for dispatch_apply in -ldispatch... " >&6; }
-if test ${ac_cv_lib_dispatch_dispatch_apply+y}
-then :
-  printf %s "(cached) " >&6
-else $as_nop
-  ac_check_lib_save_LIBS=$LIBS
-LIBS="-ldispatch  $LIBS"
-cat confdefs.h - <<_ACEOF >conftest.$ac_ext
-/* end confdefs.h.  */
-
-/* Override any GCC internal prototype to avoid an error.
-   Use char because int might match the return type of a GCC
-   builtin and then its argument prototype would still apply.  */
-char dispatch_apply ();
-int
-main (void)
-{
-return dispatch_apply ();
-  ;
-  return 0;
-}
-_ACEOF
-if ac_fn_c_try_link "$LINENO"
-then :
-  ac_cv_lib_dispatch_dispatch_apply=yes
-else $as_nop
-  ac_cv_lib_dispatch_dispatch_apply=no
-fi
-rm -f core conftest.err conftest.$ac_objext conftest.beam \
-    conftest$ac_exeext conftest.$ac_ext
-LIBS=$ac_check_lib_save_LIBS
-fi
-{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_dispatch_dispatch_apply" >&5
-printf "%s\n" "$ac_cv_lib_dispatch_dispatch_apply" >&6; }
-if test "x$ac_cv_lib_dispatch_dispatch_apply" = xyes
-then :
-  MKLDFLAGS="$MKLDFLAGS-ldispatch -lBlocksRuntime "
-
-fi
-
 fi
 # check for SSE2 support only we don't enable NEON explicitly
 if test "$neon_support" != yes; then
@@ -4590,7 +4506,7 @@ if [ -d "$cuda_prefix" ]; then
 	DEFINE_MACROS="$DEFINE_MACROS-D HAVE_CUDA "
 
 	# CUB requires stdc++, unfortunately.
-	MKLDFLAGS="$MKLDFLAGS-lcuda -lcudart -lcublas -lcurand -lstdc++ "
+	MKLDFLAGS="$MKLDFLAGS-lcudart -lcublas -lcurand -lstdc++ "
 
 	CPPFLAGS="$CPPFLAGS-I$cuda_prefix/include "
 
@@ -4742,6 +4658,90 @@ else
 
 	{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: no" >&5
 printf "%s\n" "no" >&6; }
+	# only check dispatch if compiled with clang and no cuda.
+	if test $CC = clang; then
+		 		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking dispatch/dispatch.h presence" >&5
+printf %s "checking dispatch/dispatch.h presence... " >&6; }
+if test ${ax_cv_check_cflags_dispatch_dispatch_h+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+
+		ax_check_save_flags=$CFLAGS
+		CFLAGS="$CFLAGS dispatch/dispatch.h"
+		cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+#include <dispatch/dispatch.h>
+int
+main (void)
+{
+
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_cpp "$LINENO"
+then :
+  ax_cv_check_cflags_dispatch_dispatch_h=yes
+else $as_nop
+  ax_cv_check_cflags_dispatch_dispatch_h=no
+fi
+rm -f conftest.err conftest.i conftest.$ac_ext
+		CFLAGS=$ax_check_save_flags
+fi
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ax_cv_check_cflags_dispatch_dispatch_h" >&5
+printf "%s\n" "$ax_cv_check_cflags_dispatch_dispatch_h" >&6; }
+	if test "x$ax_cv_check_cflags_dispatch_dispatch_h" = xyes
+then :
+  DEFINE_MACROS="$DEFINE_MACROS-D USE_DISPATCH "
+ MKCFLAGS="$MKCFLAGS-fblocks "
+
+else $as_nop
+  :
+fi
+
+		{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: checking for dispatch_apply in -ldispatch" >&5
+printf %s "checking for dispatch_apply in -ldispatch... " >&6; }
+if test ${ac_cv_lib_dispatch_dispatch_apply+y}
+then :
+  printf %s "(cached) " >&6
+else $as_nop
+  ac_check_lib_save_LIBS=$LIBS
+LIBS="-ldispatch  $LIBS"
+cat confdefs.h - <<_ACEOF >conftest.$ac_ext
+/* end confdefs.h.  */
+
+/* Override any GCC internal prototype to avoid an error.
+   Use char because int might match the return type of a GCC
+   builtin and then its argument prototype would still apply.  */
+char dispatch_apply ();
+int
+main (void)
+{
+return dispatch_apply ();
+  ;
+  return 0;
+}
+_ACEOF
+if ac_fn_c_try_link "$LINENO"
+then :
+  ac_cv_lib_dispatch_dispatch_apply=yes
+else $as_nop
+  ac_cv_lib_dispatch_dispatch_apply=no
+fi
+rm -f core conftest.err conftest.$ac_objext conftest.beam \
+    conftest$ac_exeext conftest.$ac_ext
+LIBS=$ac_check_lib_save_LIBS
+fi
+{ printf "%s\n" "$as_me:${as_lineno-$LINENO}: result: $ac_cv_lib_dispatch_dispatch_apply" >&5
+printf "%s\n" "$ac_cv_lib_dispatch_dispatch_apply" >&6; }
+if test "x$ac_cv_lib_dispatch_dispatch_apply" = xyes
+then :
+  MKLDFLAGS="$MKLDFLAGS-ldispatch -lBlocksRuntime "
+
+fi
+
+	fi
 fi
 
 # check for MPS support

diff --git a/lib/configure.ac b/lib/configure.ac
@@ -107,7 +107,7 @@ AC_MSG_CHECKING([fftw3])
 AC_ARG_ENABLE(fftw3, [AS_HELP_STRING([--disable-fftw3], [disable FFTW3 (GPL License)])], [fftw3_enable=$enableval], [fftw3_enable="yes"])
 if test "$fftw3_enable" != no; then
 	AX_CHECK_HEADER_PRESENCE([fftw3.h],
-		[AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D HAVE_FFTW3 "]) AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-lfftw3 -lfftw3f -lpthread "])])
+		[AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D HAVE_FFTW3 "]) AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-lfftw3 -lfftw3f "])])
 else
 	AC_MSG_RESULT([disabled])
 fi
@@ -135,12 +135,6 @@ if test -n "${OPENMP_CFLAGS}"; then
 	AC_SUBST(MKCFLAGS, ["$MKCFLAGS$OPENMP_CFLAGS "])
 	AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS$OPENMP_CFLAGS "])
 fi
-# only check dispatch if compiled with clang
-if test $CC = clang; then
-	AX_CHECK_HEADER_PRESENCE([dispatch/dispatch.h],
-		[AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D USE_DISPATCH "]) AC_SUBST(MKCFLAGS, ["$MKCFLAGS-fblocks "])])
-	AC_CHECK_LIB(dispatch, dispatch_apply, [AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-ldispatch -lBlocksRuntime "])])
-fi
 # check for SSE2 support only we don't enable NEON explicitly
 if test "$neon_support" != yes; then
 	AX_CHECK_HEADER_PRESENCE(xmmintrin.h,
@@ -165,7 +159,7 @@ AC_ARG_WITH(cuda, [AS_HELP_STRING([--with-cuda], [CUDA installation [ARG=/usr/lo
 if [[ -d "$cuda_prefix" ]]; then
 	AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D HAVE_CUDA "])
 	# CUB requires stdc++, unfortunately.
-	AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-lcuda -lcudart -lcublas -lcurand -lstdc++ "])
+	AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-lcudart -lcublas -lcurand -lstdc++ "])
 	AC_SUBST(CPPFLAGS, ["$CPPFLAGS-I$cuda_prefix/include "])
 	if [[ -d "$cuda_prefix/lib64" ]]; then
 		AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-L$cuda_prefix/lib64 "])
@@ -196,6 +190,12 @@ else
 	AC_SUBST(CUDA_COMPAT_LIB, [""])
 	AC_SUBST(CUDA_CMD_LIB, [""])
 	AC_MSG_RESULT(no)
+	# only check dispatch if compiled with clang and no cuda.
+	if test $CC = clang; then
+		AX_CHECK_HEADER_PRESENCE([dispatch/dispatch.h],
+			[AC_SUBST(DEFINE_MACROS, ["$DEFINE_MACROS-D USE_DISPATCH "]) AC_SUBST(MKCFLAGS, ["$MKCFLAGS-fblocks "])])
+		AC_CHECK_LIB(dispatch, dispatch_apply, [AC_SUBST(MKLDFLAGS, ["$MKLDFLAGS-ldispatch -lBlocksRuntime "])])
+	fi
 fi
 
 # check for MPS support

diff --git a/lib/nnc/_ccv_nnc_stream.h b/lib/nnc/_ccv_nnc_stream.h
@@ -22,6 +22,8 @@ struct ccv_nnc_stream_signal_s {
 
 typedef struct {
 	// Empty, this will hold things such as NCCL communicator in subclass.
+	// Putting dummy because C / C++ treats empty struct differently.
+	int dummy;
 } ccv_nnc_stream_resource_container_t;
 
 struct ccv_nnc_stream_context_s {

diff --git a/test/int/nnc/cudnn.tests.c b/test/int/nnc/cudnn.tests.c
@@ -5377,7 +5377,7 @@ TEST_CASE("cudnn forward convolution transpose")
 	ccv_nnc_stream_context_free(stream_context);
 	ccv_nnc_tensor_t* c = ccv_nnc_tensor_new(0, CPU_TENSOR_NHWC(32F, BATCH_SIZE, INPUT_SIZE, INPUT_SIZE, INPUT_DIM), 0);
 	ccv_nnc_cmd_exec(move, ccv_nnc_no_hint, 0, TENSOR_LIST(gc), TENSOR_LIST(c), 0);
-	REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 1e-4, "output from cudnn should match from CPU");
+	REQUIRE_ARRAY_EQ_WITH_TOLERANCE(float, b->data.f32, c->data.f32, BATCH_SIZE * INPUT_DIM * INPUT_SIZE * INPUT_SIZE, 2e-4, "output from cudnn should match from CPU");
 	ccv_nnc_tensor_free(c);
 	ccv_nnc_tensor_free(gc);
 	ccv_nnc_tensor_free(bias);