Add flag for using optimized TFLite CPU kernels on Android

Add an experimental flag which allows opting in to a set of highly
optimized floating point kernels provided via the XNNPACK delegate.
This is offered as a preview, with the plan to enable these kernels
by default in a future release. The flag can be enabled via:

  Interpreter.Options options =
      new Interpreter.Options().setUseXNNPACK(true);

See tensorflow/lite/delegates/xnnpack/README.md for more details about
these kernels and the associated delegate functionality.

PiperOrigin-RevId: 316909226
Change-Id: Ib60cf259225b8a48a9830ccbb24ec10534b038ce

tensorflow/lite/delegates/xnnpack/BUILD

@@ -21,6 +21,7 @@ cc_library(
     linkstatic = True,
     deps = [
         "//tensorflow/lite:kernel_api",
+        "//tensorflow/lite:minimal_logging",
         "//tensorflow/lite:util",
         "//tensorflow/lite/c:common",
         "//tensorflow/lite/schema:schema_fbs",
@@ -47,6 +48,7 @@ cc_library(
     linkstatic = True,
     deps = [
         "//tensorflow/lite:kernel_api",
+        "//tensorflow/lite:minimal_logging",
         "//tensorflow/lite:util",
         "//tensorflow/lite/c:common",
         "//tensorflow/lite/schema:schema_fbs",

tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc

@@ -32,6 +32,7 @@ limitations under the License.
 #include "tensorflow/lite/builtin_ops.h"
 #include "tensorflow/lite/c/builtin_op_data.h"
 #include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/minimal_logging.h"
 #include "tensorflow/lite/tools/optimize/sparsity/format_converter.h"
 
 namespace tflite {
@@ -52,6 +53,8 @@ class Delegate {
           pthreadpool_create(static_cast<size_t>(options->num_threads)));
     }
 #endif
+    TFLITE_LOG_PROD_ONCE(tflite::TFLITE_LOG_INFO,
+                         "Created TensorFlow Lite XNNPACK delegate for CPU.");
   }
 
   TfLiteIntArray* PrepareOpsToDelegate(TfLiteContext* context);

tensorflow/lite/java/BUILD

@@ -408,6 +408,7 @@ tflite_jni_binary(
         "//tensorflow/lite/c:c_api",
         "//tensorflow/lite/c:c_api_experimental",
         "//tensorflow/lite/delegates/nnapi/java/src/main/native",
+        "//tensorflow/lite/delegates/xnnpack:xnnpack_delegate",
         "//tensorflow/lite/java/src/main/native",
     ],
 )

tensorflow/lite/java/src/main/java/org/tensorflow/lite/Interpreter.java

@@ -137,10 +137,37 @@ public final class Interpreter implements AutoCloseable {
       return this;
     }
 
+    /**
+     * Experimental: Enable an optimized set of floating point CPU kernels (provided by XNNPACK).
+     *
+     * <p>Enabling this flag will enable use of a new, highly optimized set of CPU kernels provided
+     * via the XNNPACK delegate. Currently, this is restricted to a subset of floating point
+     * operations. Eventually, we plan to enable this by default, as it can provide significant
+     * peformance benefits for many classes of floating point models. See
+     * https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/delegates/xnnpack/README.md
+     * for more details.
+     *
+     * <p>Things to keep in mind when enabling this flag:
+     *
+     * <ul>
+     *   <li>Startup time and resize time may increase.
+     *   <li>Baseline memory consumption may increase.
+     *   <li>Compatibility with other delegates (e.g., GPU) has not been fully validated.
+     *   <li>Quantized models will not see any benefit.
+     * </ul>
+     *
+     * <p>WARNING: This is an experimental interface that is subject to change.
+     */
+    public Options setUseXNNPACK(boolean useXNNPACK) {
+      this.useXNNPACK = useXNNPACK;
+      return this;
+    }
+
     int numThreads = -1;
     Boolean useNNAPI;
     Boolean allowFp16PrecisionForFp32;
     Boolean allowBufferHandleOutput;
+    Boolean useXNNPACK;
     final List<Delegate> delegates = new ArrayList<>();
   }
 

tensorflow/lite/java/src/main/java/org/tensorflow/lite/NativeInterpreterWrapper.java

@@ -80,6 +80,10 @@ final class NativeInterpreterWrapper implements AutoCloseable {
       allowBufferHandleOutput(interpreterHandle, options.allowBufferHandleOutput.booleanValue());
     }
     applyDelegates(options);
+    if (options.useXNNPACK != null) {
+      useXNNPACK(
+          interpreterHandle, errorHandle, options.useXNNPACK.booleanValue(), options.numThreads);
+    }
     allocateTensors(interpreterHandle, errorHandle);
     this.isMemoryAllocated = true;
   }
@@ -438,6 +442,9 @@ final class NativeInterpreterWrapper implements AutoCloseable {
 
   private static native void allowBufferHandleOutput(long interpreterHandle, boolean allow);
 
+  private static native void useXNNPACK(
+      long interpreterHandle, long errorHandle, boolean state, int numThreads);
+
   private static native long createErrorReporter(int size);
 
   private static native long createModel(String modelPathOrBuffer, long errorHandle);

tensorflow/lite/java/src/main/native/BUILD

@@ -31,6 +31,7 @@ cc_library(
         "//tensorflow/lite:string_util",
         "//tensorflow/lite:util",
         "//tensorflow/lite/c:common",
+        "//tensorflow/lite/delegates/xnnpack:xnnpack_delegate_hdrs_only",
         "//tensorflow/lite/experimental/tflite_api_dispatcher:tflite_api_dispatcher_with_kernels",
         "//tensorflow/lite/java/jni",
     ],

tensorflow/lite/java/src/main/native/nativeinterpreterwrapper_jni.cc

@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
+#include <dlfcn.h>
 #include <jni.h>
 #include <stdio.h>
 #include <time.h>
@@ -20,6 +21,7 @@ limitations under the License.
 #include <vector>
 
 #include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
 #include "tensorflow/lite/experimental/tflite_api_dispatcher/tflite_api_dispatcher.h"
 #include "tensorflow/lite/java/src/main/native/jni_utils.h"
 #include "tensorflow/lite/util.h"
@@ -323,6 +325,59 @@ Java_org_tensorflow_lite_NativeInterpreterWrapper_allowBufferHandleOutput(
   interpreter->SetAllowBufferHandleOutput(allow);
 }
 
+JNIEXPORT void JNICALL
+Java_org_tensorflow_lite_NativeInterpreterWrapper_useXNNPACK(
+    JNIEnv* env, jclass clazz, jlong handle, jlong error_handle, jboolean state,
+    jint num_threads) {
+  // If not using xnnpack, simply don't apply the delegate.
+  if (!state) {
+    return;
+  }
+
+  tflite_api_dispatcher::Interpreter* interpreter =
+      convertLongToInterpreter(env, handle);
+  if (interpreter == nullptr) {
+    return;
+  }
+
+  BufferErrorReporter* error_reporter =
+      convertLongToErrorReporter(env, error_handle);
+  if (error_reporter == nullptr) {
+    return;
+  }
+
+  // We use dynamic loading to avoid taking a hard dependency on XNNPack.
+  // This allows clients that use trimmed builds to save on binary size.
+  auto xnnpack_options_default =
+      reinterpret_cast<decltype(TfLiteXNNPackDelegateOptionsDefault)*>(
+          dlsym(RTLD_DEFAULT, "TfLiteXNNPackDelegateOptionsDefault"));
+  auto xnnpack_create =
+      reinterpret_cast<decltype(TfLiteXNNPackDelegateCreate)*>(
+          dlsym(RTLD_DEFAULT, "TfLiteXNNPackDelegateCreate"));
+  auto xnnpack_delete =
+      reinterpret_cast<decltype(TfLiteXNNPackDelegateDelete)*>(
+          dlsym(RTLD_DEFAULT, "TfLiteXNNPackDelegateDelete"));
+
+  if (xnnpack_options_default && xnnpack_create && xnnpack_delete) {
+    TfLiteXNNPackDelegateOptions options = xnnpack_options_default();
+    if (num_threads > 0) {
+      options.num_threads = num_threads;
+    }
+    tflite_api_dispatcher::Interpreter::TfLiteDelegatePtr delegate(
+        xnnpack_create(&options), xnnpack_delete);
+    if (interpreter->ModifyGraphWithDelegate(std::move(delegate)) !=
+        kTfLiteOk) {
+      ThrowException(env, kIllegalArgumentException,
+                     "Internal error: Failed to apply XNNPACK delegate: %s",
+                     error_reporter->CachedErrorMessage());
+    }
+  } else {
+    ThrowException(env, kIllegalArgumentException,
+                   "Failed to load XNNPACK delegate from current runtime. "
+                   "Have you added the necessary dependencies?");
+  }
+}
+
 JNIEXPORT void JNICALL
 Java_org_tensorflow_lite_NativeInterpreterWrapper_numThreads(JNIEnv* env,
                                                              jclass clazz,

tensorflow/lite/java/src/test/java/org/tensorflow/lite/InterpreterMobileNetTest.java

@@ -54,6 +54,16 @@ public final class InterpreterMobileNetTest {
     runMobileNetFloatTest(new Interpreter.Options().setNumThreads(2));
   }
 
+  @Test
+  public void testMobileNetEnhancedCpuKernels() {
+    runMobileNetFloatTest(new Interpreter.Options().setUseXNNPACK(true));
+  }
+
+  @Test
+  public void testMobileNetEnhancedCpuKernelsMultithreaded() {
+    runMobileNetFloatTest(new Interpreter.Options().setUseXNNPACK(true).setNumThreads(2));
+  }
+
   @Test
   public void testMobileNetQuantized() {
     runMobileNetQuantizedTest(new Interpreter.Options());
@@ -64,6 +74,12 @@ public final class InterpreterMobileNetTest {
     runMobileNetQuantizedTest(new Interpreter.Options().setNumThreads(2));
   }
 
+  @Test
+  public void testMobileNetQuantizedEnhancedCpu() {
+    // The "enhanced CPU flag" should only impact float models, this is a sanity test to confirm.
+    runMobileNetQuantizedTest(new Interpreter.Options().setUseXNNPACK(true));
+  }
+
   private static void runMobileNetFloatTest(Interpreter.Options options) {
     ByteBuffer img =
         TestUtils.getTestImageAsFloatByteBuffer(

tensorflow/lite/java/src/test/java/org/tensorflow/lite/InterpreterTest.java

@@ -409,6 +409,38 @@ public final class InterpreterTest {
     interpreter.close();
   }
 
+  @Test
+  public void testUseXNNPACK() throws Exception {
+    Interpreter interpreter =
+        new Interpreter(MODEL_BUFFER, new Interpreter.Options().setUseXNNPACK(true));
+    float[] oneD = {1.23f, 6.54f, 7.81f};
+    float[][] twoD = {oneD, oneD, oneD, oneD, oneD, oneD, oneD, oneD};
+    float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD};
+    float[][][][] fourD = {threeD, threeD};
+    float[][][][] parsedOutputs = new float[2][8][8][3];
+    interpreter.run(fourD, parsedOutputs);
+    float[] outputOneD = parsedOutputs[0][0][0];
+    float[] expected = {3.69f, 19.62f, 23.43f};
+    assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder();
+    interpreter.close();
+  }
+
+  @Test
+  public void testResizeWithEnhancedCpuKernels() throws Exception {
+    Interpreter interpreter =
+        new Interpreter(MODEL_BUFFER, new Interpreter.Options().setUseXNNPACK(true));
+    float[] input = {1.f};
+    float[] output = new float[1];
+    interpreter.run(input, output);
+    assertThat(output).usingTolerance(0.1f).containsExactly(new float[] {3.f}).inOrder();
+
+    // The new input shape should trigger a resize. Inference should still work properly.
+    float[] input2 = {1.f, 2.f};
+    float[] output2 = new float[2];
+    interpreter.run(input2, output2);
+    assertThat(output2).usingTolerance(0.1f).containsExactly(new float[] {3.f, 6.f}).inOrder();
+  }
+
   @Test
   public void testRedundantClose() throws Exception {
     Interpreter interpreter = new Interpreter(MODEL_BUFFER);