fix build due to cupti static link.

PiperOrigin-RevId: 313880914
Change-Id: I4cf691e60fd28e773f124e440ebb30f054d5f2e9

tensorflow/core/profiler/internal/gpu/BUILD

@@ -98,11 +98,18 @@ tf_cuda_library(
     ] + tf_additional_cupti_deps(),
 )
 
+# Rationale for linkstatic: The symbols in libcupti_static.a have hidden
+# visibility. The wrapper will fail to find them if it's ever built as a
+# shared library. This is the same issue as b/11094727. Always linking
+# the wrapper statically works around the issue. An alternative would be
+# to patch libcupti_static, but it's not worth the trouble considering
+# that the wrapper is about the only direct user.
 tf_cuda_library(
     name = "cupti_wrapper",
     srcs = if_cuda_is_configured_compat(["cupti_wrapper.cc"]),
     hdrs = if_cuda_is_configured_compat(["cupti_wrapper.h"]),
     copts = tf_copts(),
+    linkstatic = 1,
     visibility = ["//visibility:public"],
     deps = [
         ":cupti_interface",

tensorflow/stream_executor/cuda/cupti_stub.cc

@@ -14,6 +14,7 @@ limitations under the License.
 ==============================================================================*/
 
 #include "third_party/gpus/cuda/extras/CUPTI/include/cupti.h"
+#include "third_party/gpus/cuda/include/cuda.h"
 // IWYU pragma: no_include "perftools/gputools/executor/stream_executor.h"
 #include "tensorflow/stream_executor/lib/env.h"
 #include "tensorflow/stream_executor/platform/dso_loader.h"