Avoid ToString() in Eager's TFE_Execute.

Also use InlinedVector instead of std::vector for non-async path

Before:
Benchmark              Time(ns)        CPU(ns)     Iterations
-------------------------------------------------------------
BM_Execute/0               1895           1898         360200  Execute
BM_Execute/1               1193           1942         358322  ExecuteAsync
BM_ExecuteFunction/0       5812           5825         100000  ExecuteFunction
BM_ExecuteFunction/1       5015           5374         100000  ExecuteFunctionAsync

After:
Benchmark              Time(ns)        CPU(ns)     Iterations
-------------------------------------------------------------
BM_Execute/0               1604           1607         428262  Execute
BM_Execute/1               1150           1765         404821  ExecuteAsync
BM_ExecuteFunction/0       5615           5626         100000  ExecuteFunction
BM_ExecuteFunction/1       5111           5476         100000  ExecuteFunctionAsync
PiperOrigin-RevId: 193218331

2 files changed, 7 insertions, 16 deletions

diff --git a/tensorflow/c/eager/c_api.cc b/tensorflow/c/eager/c_api.cc
index c96a38dec3..393851d13c 100644
--- a/tensorflow/c/eager/c_api.cc
+++ b/tensorflow/c/eager/c_api.cc
@@ -116,9 +116,7 @@ TFE_Context* TFE_NewContext(const TFE_ContextOptions* opts, TF_Status* status) {
                          opts->async, std::move(device_mgr), r);
 }
 
-void TFE_DeleteContext(TFE_Context* ctx, TF_Status* status) {
-  delete ctx;
-}
+void TFE_DeleteContext(TFE_Context* ctx, TF_Status* status) { delete ctx; }
 
 TF_DeviceList* TFE_ContextListDevices(TFE_Context* ctx, TF_Status* status) {
   TF_DeviceList* list = new TF_DeviceList;
@@ -581,7 +579,6 @@ tensorflow::Device* SelectDevice(const tensorflow::NodeDef& ndef,
   return nullptr;
 }
 
-
 #ifdef TENSORFLOW_EAGER_USE_XLA
 // Synthesizes and returns a wrapper function over `op`, which must be a
 // primitive op (e.g. matmul).
@@ -725,9 +722,7 @@ std::unique_ptr<TFE_Op> BuildXlaLaunch(TFE_Op* op, TF_Status* status) {
   }
 
   const tensorflow::FunctionDef* fdef;
-  {
-    fdef = op->ctx->context.FindFunctionDef(op->name);
-  }
+  { fdef = op->ctx->context.FindFunctionDef(op->name); }
   std::vector<TF_DataType> const_input_types;
   std::vector<TF_DataType> arg_input_types;
   tensorflow::gtl::FlatMap<int, int> op_input_to_func_input;
@@ -940,8 +935,8 @@ void TFE_Execute(TFE_Op* op, TFE_TensorHandle** retvals, int* num_retvals,
   } else {
     // Execute checks if retvals[i] is nullptr or not to figure if it needs to
     // allocate it.
-    std::vector<tensorflow::TensorHandle*> handle_retvals(*num_retvals,
-                                                          nullptr);
+    tensorflow::gtl::InlinedVector<tensorflow::TensorHandle*, 2> handle_retvals(
+        *num_retvals);
     status->status = tensorflow::EagerExecute(
         &op->ctx->context, op->device, op->inputs, kernel, maybe_stats.get(),
         handle_retvals.data(), *num_retvals);
@@ -1091,7 +1086,6 @@ void SetOpAttrValueScalar(TFE_Context* ctx, TFE_Op* op,
 }
 }  // namespace tensorflow
 
-
 TFE_Op::~TFE_Op() {
   for (tensorflow::TensorHandle* h : inputs) {
     h->Unref();
diff --git a/tensorflow/c/eager/runtime.cc b/tensorflow/c/eager/runtime.cc
index abe2793ce8..e6c51ab17a 100644
--- a/tensorflow/c/eager/runtime.cc
+++ b/tensorflow/c/eager/runtime.cc
@@ -184,8 +184,7 @@ void CombineUnordered(const tensorflow::Fprint128& a,
 
 inline tensorflow::Fprint128 CacheKeyHelper(StringPiece s,
                                             const tensorflow::Fprint128& b) {
-  // TODO(agarwal): avoid ToString().
-  tensorflow::Fprint128 a = tensorflow::Fingerprint128(s.ToString());
+  tensorflow::Fprint128 a = tensorflow::Fingerprint128(s);
   return FingerprintCat128(a, b);
 }
 
@@ -213,10 +212,8 @@ tensorflow::Fprint128 AttrBuilder::CacheKey(const string& device) const {
     if (node_def_finalized_) return f;
   }
   for (const auto& p : string_attrs_) {
-    // TODO(agarwal): avoid ToString().
-    CombineUnordered(CacheKeyHelper(p.first, tensorflow::Fingerprint128(
-                                                 p.second.ToString())),
-                     &f);
+    CombineUnordered(
+        CacheKeyHelper(p.first, tensorflow::Fingerprint128(p.second)), &f);
   }
   for (const auto& p : int_attrs_) {
     CombineUnordered(CacheKeyHelper(p.first, static_cast<uint64>(p.second)),