Add BF16 Softmax/SoftmaxGrad and fix accuracy issue by accum type

tensorflow/c/BUILD

@@ -334,6 +334,9 @@ tf_cuda_library(
         ":checkpoint_reader",
         "//tensorflow/c/eager:c_api",
         "//tensorflow/c/eager:c_api_internal",
+        "//tensorflow/c/eager:tfe_context_internal",
+        "//tensorflow/c/eager:tfe_op_internal",
+        "//tensorflow/c/eager:tfe_tensorhandle_internal",
         "//tensorflow/compiler/jit:flags",
         "//tensorflow/core:core_cpu",
         "//tensorflow/core:framework",
@@ -729,3 +732,11 @@ tf_cuda_library(
     ],
     alwayslink = 1,
 )
+
+cc_library(
+    name = "conversion_macros",
+    hdrs = [
+        "conversion_macros.h",
+    ],
+    visibility = ["//tensorflow:__subpackages__"],
+)

tensorflow/c/c_api_experimental.cc

@@ -21,6 +21,9 @@ limitations under the License.
 #include "tensorflow/c/checkpoint_reader.h"
 #include "tensorflow/c/eager/c_api.h"
 #include "tensorflow/c/eager/c_api_internal.h"
+#include "tensorflow/c/eager/tfe_context_internal.h"
+#include "tensorflow/c/eager/tfe_op_internal.h"
+#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"
 #include "tensorflow/compiler/jit/flags.h"
 #include "tensorflow/core/common_runtime/eager/attr_builder.h"
 #include "tensorflow/core/common_runtime/eager/context.h"
@@ -686,8 +689,7 @@ TFE_TensorHandle* TFE_NewTensorHandleFromScalar(TF_DataType data_type,
   std::memcpy(tensorflow::TensorCApi::Buffer(tensor)->data(), data, len);
 
   status->status = tensorflow::Status::OK();
-  return new TFE_TensorHandle{
-      tensorflow::TensorHandle::CreateLocalHandle(tensor)};
+  return tensorflow::wrap(tensorflow::TensorHandle::CreateLocalHandle(tensor));
 }
 
 namespace {
@@ -708,7 +710,7 @@ tensorflow::Status EnableCollectiveOps(const tensorflow::ServerDef& server_def,
 
   // New server created for new server_def. Unused if updating server_def.
   tensorflow::EagerContext* context =
-      tensorflow::ContextFromInterface(ctx->context);
+      tensorflow::ContextFromInterface(tensorflow::unwrap(ctx));
   tensorflow::GrpcServer* grpc_server =
       dynamic_cast<tensorflow::GrpcServer*>(context->GetServer());
   if (grpc_server == nullptr) {
@@ -822,14 +824,13 @@ void TFE_InferShapes(TFE_Op* tfe_op, TF_ShapeAndTypeList* input_shapes,
 
   const int num_inputs = input_shapes->num_items;
   NodeDef node_def;
-  node_def.set_name(tfe_op->operation->Name());
-  node_def.set_op(tfe_op->operation->Name());
+  tensorflow::AbstractOperationInterface* op = tensorflow::unwrap(tfe_op);
+  node_def.set_name(op->Name());
+  node_def.set_op(op->Name());
   for (int i = 0; i < num_inputs; ++i) {
     node_def.add_input("dummy_input");
   }
-  OperationFromInterface(tfe_op->operation)
-      ->Attrs()
-      .FillAttrValueMap(node_def.mutable_attr());
+  OperationFromInterface(op)->Attrs().FillAttrValueMap(node_def.mutable_attr());
 
   const tensorflow::OpRegistrationData* op_reg_data;
   status->status =

tensorflow/c/conversion_macros.h

@@ -0,0 +1,30 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_C_CONVERSION_MACROS_H_
+#define TENSORFLOW_C_CONVERSION_MACROS_H_
+
+#define DEFINE_CONVERSION_FUNCTIONS(cpp_impl, wrapper) \
+  inline cpp_impl *unwrap(wrapper *w) {                \
+    return reinterpret_cast<cpp_impl *>(w);            \
+  }                                                    \
+                                                       \
+  inline const cpp_impl *unwrap(const wrapper *w) {    \
+    return reinterpret_cast<const cpp_impl *>(w);      \
+  }                                                    \
+                                                       \
+  inline wrapper *wrap(cpp_impl *i) { return reinterpret_cast<wrapper *>(i); }
+
+#endif  // TENSORFLOW_C_CONVERSION_MACROS_H_

tensorflow/c/eager/BUILD

@@ -50,7 +50,6 @@ tf_cuda_library(
             ":tfe_tensor_debug_info_internal",
             ":tfe_tensorhandle_internal",
             "@com_google_absl//absl/algorithm:container",
-            "@com_google_absl//absl/container:fixed_array",
             "@com_google_absl//absl/types:span",
             "@com_google_absl//absl/types:variant",
             "//tensorflow/c:c_api",
@@ -110,13 +109,10 @@ filegroup(
         "operation_interface.h",
         "tensor_handle_interface.h",
         "tfe_cancellation_manager_internal.h",
-        "tfe_context_internal.h",
         "tfe_executor_internal.h",
         "tfe_monitoring_internal.h",
         "tfe_op_attrs_internal.h",
-        "tfe_op_internal.h",
         "tfe_tensor_debug_info_internal.h",
-        "tfe_tensorhandle_internal.h",
     ],
     visibility = [
         "//tensorflow/core:__pkg__",
@@ -205,6 +201,7 @@ cc_library(
     ],
     deps = [
         ":context_interface",
+        "//tensorflow/c:conversion_macros",
     ],
 )
 
@@ -249,8 +246,6 @@ cc_library(
         "//tensorflow:internal",
     ],
     deps = [
-        ":tfe_context_internal",
-        ":tfe_op_internal",
         "//tensorflow/c:tf_status",
         "//tensorflow/core:protos_all_cc",
         "//tensorflow/core/common_runtime/eager:attr_builder",
@@ -265,6 +260,7 @@ cc_library(
     ],
     deps = [
         ":operation_interface",
+        "//tensorflow/c:conversion_macros",
     ],
 )
 
@@ -287,6 +283,7 @@ cc_library(
     ],
     deps = [
         ":tensor_handle_interface",
+        "//tensorflow/c:conversion_macros",
     ],
 )
 
@@ -327,6 +324,8 @@ tf_cuda_cc_test(
         ":c_api_experimental",
         ":c_api_internal",
         ":c_api_test_util",
+        ":tfe_op_internal",
+        ":tfe_tensorhandle_internal",
         "//tensorflow/c:c_test_util",
         "//tensorflow/core:lib",
         "//tensorflow/core:lib_internal",
@@ -351,6 +350,7 @@ tf_cuda_cc_test(
         ":c_api_experimental",
         ":c_api_internal",
         ":c_api_test_util",
+        ":tfe_tensorhandle_internal",
         "//tensorflow/c:c_test_util",
         "//tensorflow/core:lib",
         "//tensorflow/core:protos_all_cc",
@@ -384,6 +384,9 @@ tf_cuda_library(
         "//conditions:default": [
             ":c_api",
             ":c_api_internal",
+            ":tfe_context_internal",
+            ":tfe_op_internal",
+            ":tfe_tensorhandle_internal",
             "//tensorflow/c:c_api",
             "//tensorflow/c:c_api_internal",
             "//tensorflow/core:core_cpu",
@@ -548,8 +551,9 @@ cc_library(
     deps = [
         ":c_api",
         ":c_api_experimental",
-        ":c_api_internal",
+        ":tfe_tensorhandle_internal",
         "//tensorflow/c:tf_status_helper",
+        "//tensorflow/c:tf_status_internal",
         "//tensorflow/core:framework",
         "//tensorflow/core:framework_internal",
         "//tensorflow/core:lib",