Mark Interpreter:UseNNAPI(bool) deprecated

Prefer using the NnApiDelegate() API directly. The current API is
unreliable and causes issues with the first inference execution.

PiperOrigin-RevId: 325496322
Change-Id: I44c8fc04bcd08ce5b92e22cd170e075c0abbaecf

Author: jdduke

RELEASE.md

@@ -111,6 +111,8 @@
       string to be joined is empty.
     * `TFLiteConverter`:
       * Support optional flags `inference_input_type` and `inference_output_type` for full integer quantized models. This allows users to modify the model input and output type to integer types (`tf.int8`, `tf.uint8`) instead of defaulting to float type (`tf.float32`).
+    * Deprecate `Interpreter::UseNNAPI(bool)` C++ API
+      * Prefer using `NnApiDelegate()` and related delegate configuration methods directly.
     * <ADD RELEASE NOTES HERE>
 *   `tf.random`:
     * <ADD RELEASE NOTES HERE>

tensorflow/lite/c/BUILD

@@ -62,6 +62,7 @@ cc_library(
         "//tensorflow/lite:framework",
         "//tensorflow/lite:version",
         "//tensorflow/lite/core/api",
+        "//tensorflow/lite/delegates/nnapi:nnapi_delegate",
         "//tensorflow/lite/kernels:builtin_ops",
     ],
     alwayslink = 1,

tensorflow/lite/c/c_api.cc

@@ -17,6 +17,7 @@ limitations under the License.
 #include <memory>
 
 #include "tensorflow/lite/c/c_api_internal.h"
+#include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h"
 #include "tensorflow/lite/error_reporter.h"
 #include "tensorflow/lite/interpreter.h"
 #include "tensorflow/lite/kernels/register.h"
@@ -123,13 +124,18 @@ TfLiteInterpreter* TfLiteInterpreterCreate(
   }
 
   if (optional_options) {
-    interpreter->UseNNAPI(optional_options->use_nnapi);
-
     if (optional_options->num_threads !=
         TfLiteInterpreterOptions::kDefaultNumThreads) {
       interpreter->SetNumThreads(optional_options->num_threads);
     }
 
+    if (optional_options->use_nnapi) {
+      if (interpreter->ModifyGraphWithDelegate(tflite::NnApiDelegate()) !=
+          kTfLiteOk) {
+        return nullptr;
+      }
+    }
+
     for (auto* delegate : optional_options->delegates) {
       if (interpreter->ModifyGraphWithDelegate(delegate) != kTfLiteOk) {
         return nullptr;

tensorflow/lite/delegates/gpu/common/testing/tflite_model_reader.cc

@@ -79,7 +79,6 @@ absl::Status BuildFromFlatBuffer(const tflite::FlatBufferModel& flatbuffer,
   if (interpreter_builder(&interpreter) != kTfLiteOk || !interpreter) {
     return absl::InternalError("Unable to prepare TfLite interpreter.");
   }
-  interpreter->UseNNAPI(false);
   TfLiteDelegate delegate;
   delegate.data_ = graph;
   delegate.flags = kTfLiteDelegateFlagsNone;

tensorflow/lite/interpreter.cc

@@ -300,7 +300,12 @@ TfLiteStatus Interpreter::SetExecutionPlan(const std::vector<int>& new_plan) {
   return primary_subgraph().SetExecutionPlan(new_plan);
 }
 
-void Interpreter::UseNNAPI(bool enable) { primary_subgraph().UseNNAPI(enable); }
+void Interpreter::UseNNAPI(bool enable) {
+  TFLITE_LOG_PROD_ONCE(TFLITE_LOG_INFO,
+                       "Interpreter::UseNNAPI() is deprecated. Use "
+                       "tflite::NnApiDelegate() directly instead.");
+  primary_subgraph().UseNNAPI(enable);
+}
 
 TfLiteStatus Interpreter::SetNumThreads(int num_threads) {
   if (num_threads < -1) {

tensorflow/lite/interpreter.h

@@ -365,6 +365,8 @@ class Interpreter {
   TfLiteStatus Invoke();
 
   /// Enable or disable the NN API (true to enable)
+  /// NOTE: This API is deprecated, prefer using the NNAPI delegate directly.
+  /// This method will be removed in a future release.
   void UseNNAPI(bool enable);
 
   /// Set the number of threads available to the interpreter.