Implement event based CUDA profiling. Continue to use CUPTI callback API to generate events, but avoid activity API.

PiperOrigin-RevId: 238427551

tensorflow/core/platform/default/build_config.bzl

@@ -577,7 +577,14 @@ def tf_protos_grappler():
     )
 
 def tf_additional_cupti_wrapper_deps():
-    return ["//tensorflow/core/platform/default/gpu:cupti_wrapper"]
+    return [
+        "//tensorflow/core/platform/default/gpu:cupti_wrapper",
+        "@com_google_absl//absl/base",
+        "@com_google_absl//absl/strings",
+        "@com_google_absl//absl/strings:str_format",
+        "@com_google_absl//absl/container:node_hash_map",
+        "@com_google_absl//absl/container:flat_hash_map",
+    ]
 
 def tf_additional_device_tracer_srcs():
     return ["platform/default/device_tracer.cc"]

tensorflow/core/platform/default/gpu/cupti_wrapper.cc

@@ -61,6 +61,7 @@ LIBCUPTI_WRAP(cuptiEnableCallback);
 LIBCUPTI_WRAP(cuptiEnableDomain);
 LIBCUPTI_WRAP(cuptiSubscribe);
 LIBCUPTI_WRAP(cuptiUnsubscribe);
+LIBCUPTI_WRAP(cuptiGetResultString);
 
 }  // namespace dynload
 
@@ -124,6 +125,11 @@ CUptiResult CuptiWrapper::Unsubscribe(CUpti_SubscriberHandle subscriber) {
   return dynload::cuptiUnsubscribe(subscriber);
 }
 
+CUptiResult CuptiWrapper::GetResultString(CUptiResult result,
+                                          const char** str) {
+  return dynload::cuptiGetResultString(result, str);
+}
+
 }  // namespace profiler
 }  // namespace gputools
 }  // namespace perftools

tensorflow/core/platform/default/gpu/cupti_wrapper.h

@@ -68,6 +68,8 @@ class CuptiWrapper {
                         CUpti_CallbackFunc callback, void* userdata);
 
   CUptiResult Unsubscribe(CUpti_SubscriberHandle subscriber);
+
+  CUptiResult GetResultString(CUptiResult result, const char** str);
 };
 
 }  // namespace profiler