Not yet public

PiperOrigin-RevId: 319040214
Change-Id: I95edff062ee32c713df97752b29e7172bfe6aec3

tensorflow/lite/experimental/acceleration/configuration/configuration.proto

@@ -69,6 +69,13 @@ enum NNAPIExecutionPreference {
   NNAPI_SUSTAINED_SPEED = 3;
 }
 
+enum NNAPIExecutionPriority {
+  NNAPI_PRIORITY_UNDEFINED = 0;
+  NNAPI_PRIORITY_LOW = 1;
+  NNAPI_PRIORITY_MEDIUM = 2;
+  NNAPI_PRIORITY_HIGH = 3;
+}
+
 // One possible acceleration configuration.
 message ComputeSettings {
   // Which preference to use this accelerator for.
@@ -110,6 +117,8 @@ message NNAPISettings {
   // performs less well than the TfLite built-in kernels; but allowing allows a
   // model to be partially accelerated which may be a win.
   optional bool allow_nnapi_cpu_on_android_10_plus = 7;
+
+  optional NNAPIExecutionPriority execution_priority = 8;
 }
 
 // Which GPU backend to select. Default behaviour on Android is to try OpenCL

tensorflow/lite/experimental/acceleration/configuration/nnapi_plugin.cc

@@ -39,6 +39,20 @@ ConvertExecutionPrefence(
   }
 }
 
+inline int ConvertExecutionPriority(
+    NNAPIExecutionPriority from_compatibility_priority) {
+  switch (from_compatibility_priority) {
+    case NNAPIExecutionPriority_NNAPI_PRIORITY_LOW:
+      return ANEURALNETWORKS_PRIORITY_LOW;
+    case NNAPIExecutionPriority_NNAPI_PRIORITY_MEDIUM:
+      return ANEURALNETWORKS_PRIORITY_MEDIUM;
+    case NNAPIExecutionPriority_NNAPI_PRIORITY_HIGH:
+      return ANEURALNETWORKS_PRIORITY_HIGH;
+    default:
+      return ANEURALNETWORKS_PRIORITY_DEFAULT;
+  }
+}
+
 class NnapiPlugin : public DelegatePluginInterface {
  public:
   TfLiteDelegatePtr Create() override {
@@ -80,6 +94,8 @@ class NnapiPlugin : public DelegatePluginInterface {
         ConvertExecutionPrefence(nnapi_settings->execution_preference());
     options_.disallow_nnapi_cpu =
         !nnapi_settings->allow_nnapi_cpu_on_android_10_plus();
+    options_.execution_priority =
+        ConvertExecutionPriority(nnapi_settings->execution_priority());
   }
 
  private:

tensorflow/lite/experimental/acceleration/configuration/nnapi_plugin_test.cc

@@ -18,6 +18,7 @@ limitations under the License.
 #include "flatbuffers/flatbuffers.h"  // from @flatbuffers
 #include "tensorflow/lite/c/common.h"
 #include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h"
+#include "tensorflow/lite/delegates/nnapi/nnapi_delegate_kernel.h"
 #include "tensorflow/lite/delegates/nnapi/nnapi_delegate_mock_test.h"
 #include "tensorflow/lite/experimental/acceleration/configuration/configuration_generated.h"
 #include "tensorflow/lite/experimental/acceleration/configuration/delegate_registry.h"
@@ -82,6 +83,25 @@ class NNAPIPluginTest : public ::testing::Test {
               kTfLiteOk)
         << " given input: " << input << " expected output: " << output;
   }
+  template <NNAPIExecutionPriority input, int output>
+  void CheckExecutionPriority() {
+    // Note - this uses a template since the NNAPI functions are C function
+    // pointers rather than lambdas so can't capture variables.
+    nnapi_->ANeuralNetworksCompilation_setPriority =
+        [](ANeuralNetworksCompilation* compilation, int32_t priority) {
+          return priority - output;
+        };
+    CreateDelegate(CreateNNAPISettings(fbb_, 0, 0, 0,
+                                       NNAPIExecutionPreference_UNDEFINED, 0, 0,
+                                       /*allow CPU=*/true, input));
+    // Since delegation succeeds, the model becomes immutable and hence can't
+    // reuse it.
+    SingleAddOpModel model;
+    model.Build();
+    EXPECT_EQ(model.Interpreter()->ModifyGraphWithDelegate(delegate_.get()),
+              kTfLiteOk)
+        << " given input: " << input << " expected output: " << output;
+  }
 
   void CreateDelegate(flatbuffers::Offset<NNAPISettings> settings) {
     settings_ = flatbuffers::GetTemporaryPointer(
@@ -124,6 +144,19 @@ TEST_F(NNAPIPluginTest, PassesExecutionPreference) {
                            StatefulNnApiDelegate::Options::kSustainedSpeed>();
 }
 
+TEST_F(NNAPIPluginTest, PassesExecutionPriority) {
+  nnapi_->android_sdk_version =
+      tflite::delegate::nnapi::kMinSdkVersionForNNAPI13;
+  CheckExecutionPriority<NNAPIExecutionPriority_NNAPI_PRIORITY_UNDEFINED,
+                         ANEURALNETWORKS_PRIORITY_DEFAULT>();
+  CheckExecutionPriority<NNAPIExecutionPriority_NNAPI_PRIORITY_LOW,
+                         ANEURALNETWORKS_PRIORITY_LOW>();
+  CheckExecutionPriority<NNAPIExecutionPriority_NNAPI_PRIORITY_MEDIUM,
+                         ANEURALNETWORKS_PRIORITY_MEDIUM>();
+  CheckExecutionPriority<NNAPIExecutionPriority_NNAPI_PRIORITY_HIGH,
+                         ANEURALNETWORKS_PRIORITY_HIGH>();
+}
+
 TEST_F(NNAPIPluginTest, PassesCachingParameters) {
   nnapi_->ANeuralNetworksCompilation_setCaching =
       [](ANeuralNetworksCompilation* compilation, const char* cacheDir,