Support LEAKY_RELU operator in XNNPACK delegate

PiperOrigin-RevId: 315833208
Change-Id: Icd123c1548febc03056f9ad42039e83a8a77d19f

tensorflow/lite/delegates/xnnpack/BUILD

@@ -124,6 +124,22 @@ cc_library(
     ],
 )
 
+cc_library(
+    name = "leaky_relu_tester",
+    testonly = 1,
+    srcs = ["leaky_relu_tester.cc"],
+    hdrs = ["leaky_relu_tester.h"],
+    deps = [
+        "//tensorflow/lite:framework",
+        "//tensorflow/lite:schema_fbs_version",
+        "//tensorflow/lite/c:common",
+        "//tensorflow/lite/kernels:builtin_ops",
+        "//tensorflow/lite/schema:schema_fbs",
+        "@com_google_googletest//:gtest",
+        "@flatbuffers",
+    ],
+)
+
 cc_library(
     name = "pad_tester",
     testonly = 1,
@@ -353,6 +369,21 @@ cc_test(
     ],
 )
 
+cc_test(
+    name = "leaky_relu_test",
+    srcs = ["leaky_relu_test.cc"],
+    linkopts = select({
+        "//tensorflow:emscripten": EMSCRIPTEN_LINKOPTS,
+        "//conditions:default": [],
+    }),
+    deps = [
+        ":leaky_relu_tester",
+        ":test_main",
+        ":xnnpack_delegate_test_mode",
+        "@com_google_googletest//:gtest",
+    ],
+)
+
 cc_test(
     name = "logistic_test",
     srcs = ["logistic_test.cc"],

tensorflow/lite/delegates/xnnpack/README.md

@@ -146,6 +146,10 @@ Below is the list of current operators and limitations:
 
 * Inputs and outputs must be in 32-bit floating-point format.
 
+### `LEAKY_RELU`
+
+* Inputs and outputs must be in 32-bit floating-point format.
+
 ### `LOGISTIC`
 
 * Inputs and outputs must be in 32-bit floating-point format.

tensorflow/lite/delegates/xnnpack/leaky_relu_test.cc

@@ -0,0 +1,137 @@
+/* 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.
+==============================================================================*/
+
+#include <cstdint>
+#include <functional>
+#include <memory>
+#include <random>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h"
+#include "tensorflow/lite/delegates/xnnpack/xnnpack_delegate.h"
+
+namespace tflite {
+namespace xnnpack {
+
+TEST(LeakyRelu, 4D) {
+  std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  LeakyReluTester()
+      .Shape({batch, height, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(LeakyRelu, 3D) {
+  std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  LeakyReluTester()
+      .Shape({batch, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(LeakyRelu, 2D) {
+  std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto channels = shape_rng();
+
+  LeakyReluTester().Shape({batch, channels}).Test(xnnpack_delegate.get());
+}
+
+TEST(LeakyRelu, 1D) {
+  std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+
+  LeakyReluTester().Shape({batch}).Test(xnnpack_delegate.get());
+}
+
+TEST(LeakyRelu, NegativeSlope) {
+  std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(nullptr),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  LeakyReluTester()
+      .Shape({batch, height, width, channels})
+      .NegativeSlope(-0.75f)
+      .Test(xnnpack_delegate.get());
+}
+
+TEST(LeakyRelu, MultiThreading) {
+  TfLiteXNNPackDelegateOptions delegate_options =
+      TfLiteXNNPackDelegateOptionsDefault();
+  delegate_options.num_threads = 2;
+  std::unique_ptr<TfLiteDelegate, decltype(&TfLiteXNNPackDelegateDelete)>
+      xnnpack_delegate(TfLiteXNNPackDelegateCreate(&delegate_options),
+                       TfLiteXNNPackDelegateDelete);
+
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto shape_rng =
+      std::bind(std::uniform_int_distribution<int32_t>(2, 5), std::ref(rng));
+  const auto batch = shape_rng();
+  const auto height = shape_rng();
+  const auto width = shape_rng();
+  const auto channels = shape_rng();
+
+  LeakyReluTester()
+      .Shape({batch, height, width, channels})
+      .Test(xnnpack_delegate.get());
+}
+
+}  // namespace xnnpack
+}  // namespace tflite

tensorflow/lite/delegates/xnnpack/leaky_relu_tester.cc

@@ -0,0 +1,154 @@
+/* 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.
+==============================================================================*/
+
+#include "tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h"
+
+#include <array>
+#include <cstdint>
+#include <functional>
+#include <numeric>
+#include <random>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "flatbuffers/flatbuffers.h"  // from @flatbuffers
+#include "tensorflow/lite/interpreter.h"
+#include "tensorflow/lite/kernels/register.h"
+#include "tensorflow/lite/model.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+#include "tensorflow/lite/version.h"
+
+namespace tflite {
+namespace xnnpack {
+
+void LeakyReluTester::Test(TfLiteDelegate* delegate) const {
+  std::random_device random_device;
+  auto rng = std::mt19937(random_device());
+  auto input_rng = std::bind(
+      std::uniform_real_distribution<float>(-15.0f, 15.0f), std::ref(rng));
+
+  std::vector<char> buffer = CreateTfLiteModel();
+  const Model* model = GetModel(buffer.data());
+
+  std::unique_ptr<Interpreter> delegate_interpreter;
+  ASSERT_EQ(
+      InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+          &delegate_interpreter),
+      kTfLiteOk);
+  std::unique_ptr<Interpreter> default_interpreter;
+  ASSERT_EQ(
+      InterpreterBuilder(model, ::tflite::ops::builtin::BuiltinOpResolver())(
+          &default_interpreter),
+      kTfLiteOk);
+
+  ASSERT_TRUE(delegate_interpreter);
+  ASSERT_TRUE(default_interpreter);
+
+  ASSERT_EQ(delegate_interpreter->inputs().size(), 1);
+  ASSERT_EQ(default_interpreter->inputs().size(), 1);
+
+  ASSERT_EQ(delegate_interpreter->outputs().size(), 1);
+  ASSERT_EQ(default_interpreter->outputs().size(), 1);
+
+  ASSERT_EQ(delegate_interpreter->AllocateTensors(), kTfLiteOk);
+  ASSERT_EQ(default_interpreter->AllocateTensors(), kTfLiteOk);
+
+  ASSERT_EQ(delegate_interpreter->ModifyGraphWithDelegate(delegate), kTfLiteOk);
+
+  float* default_input_data = default_interpreter->typed_tensor<float>(
+      default_interpreter->inputs()[0]);
+  std::generate(default_input_data, default_input_data + Size(),
+                std::ref(input_rng));
+
+  float* delegate_input_data = delegate_interpreter->typed_tensor<float>(
+      delegate_interpreter->inputs()[0]);
+  std::copy(default_input_data, default_input_data + Size(),
+            delegate_input_data);
+
+  ASSERT_EQ(default_interpreter->Invoke(), kTfLiteOk);
+  ASSERT_EQ(delegate_interpreter->Invoke(), kTfLiteOk);
+
+  float* default_output_data = default_interpreter->typed_tensor<float>(
+      default_interpreter->outputs()[0]);
+  float* delegate_output_data = delegate_interpreter->typed_tensor<float>(
+      delegate_interpreter->outputs()[0]);
+
+  for (size_t i = 0; i < Size(); i++) {
+    ASSERT_EQ(default_output_data[i], delegate_output_data[i]);
+  }
+}
+
+std::vector<char> LeakyReluTester::CreateTfLiteModel() const {
+  flatbuffers::FlatBufferBuilder builder;
+  flatbuffers::Offset<OperatorCode> operator_code =
+      CreateOperatorCode(builder, BuiltinOperator_LEAKY_RELU);
+
+  const std::array<flatbuffers::Offset<Buffer>, 1> buffers{{
+      CreateBuffer(builder, builder.CreateVector({})),
+  }};
+
+  const std::array<flatbuffers::Offset<Tensor>, 2> tensors{{
+      CreateTensor(
+          builder,
+          builder.CreateVector<int32_t>(Shape().data(), Shape().size()),
+          TensorType_FLOAT32),
+      CreateTensor(
+          builder,
+          builder.CreateVector<int32_t>(Shape().data(), Shape().size()),
+          TensorType_FLOAT32),
+  }};
+
+  const flatbuffers::Offset<LeakyReluOptions> leaky_relu_options =
+      CreateLeakyReluOptions(builder, NegativeSlope());
+
+  const std::array<int32_t, 1> op_inputs{{0}};
+  const std::array<int32_t, 1> op_outputs{{1}};
+  flatbuffers::Offset<Operator> op = CreateOperator(
+      builder, /*opcode_index=*/0,
+      builder.CreateVector<int32_t>(op_inputs.data(), op_inputs.size()),
+      builder.CreateVector<int32_t>(op_outputs.data(), op_outputs.size()),
+      BuiltinOptions_LeakyReluOptions, leaky_relu_options.Union());
+
+  const std::array<int32_t, 1> subgraph_inputs{{0}};
+  const std::array<int32_t, 1> subgraph_outputs{{1}};
+  flatbuffers::Offset<SubGraph> subgraph = CreateSubGraph(
+      builder, builder.CreateVector(tensors.data(), tensors.size()),
+      builder.CreateVector<int32_t>(subgraph_inputs.data(),
+                                    subgraph_inputs.size()),
+      builder.CreateVector<int32_t>(subgraph_outputs.data(),
+                                    subgraph_outputs.size()),
+      builder.CreateVector(&op, 1));
+
+  flatbuffers::Offset<flatbuffers::String> description =
+      builder.CreateString("Leaky ReLU model");
+
+  flatbuffers::Offset<Model> model_buffer = CreateModel(
+      builder, TFLITE_SCHEMA_VERSION, builder.CreateVector(&operator_code, 1),
+      builder.CreateVector(&subgraph, 1), description,
+      builder.CreateVector(buffers.data(), buffers.size()));
+
+  builder.Finish(model_buffer);
+
+  return std::vector<char>(builder.GetBufferPointer(),
+                           builder.GetBufferPointer() + builder.GetSize());
+}
+
+int32_t LeakyReluTester::ComputeSize(const std::vector<int32_t>& shape) {
+  return std::accumulate(shape.cbegin(), shape.cend(), 1,
+                         std::multiplies<int32_t>());
+}
+
+}  // namespace xnnpack
+}  // namespace tflite

tensorflow/lite/delegates/xnnpack/leaky_relu_tester.h

@@ -0,0 +1,70 @@
+/* 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_LITE_DELEGATES_XNNPACK_LEAKY_RELU_TESTER_H_
+#define TENSORFLOW_LITE_DELEGATES_XNNPACK_LEAKY_RELU_TESTER_H_
+
+#include <cstdint>
+#include <vector>
+
+#include <gtest/gtest.h>
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+
+namespace tflite {
+namespace xnnpack {
+
+class LeakyReluTester {
+ public:
+  LeakyReluTester() = default;
+  LeakyReluTester(const LeakyReluTester&) = delete;
+  LeakyReluTester& operator=(const LeakyReluTester&) = delete;
+
+  inline LeakyReluTester& Shape(std::initializer_list<int32_t> shape) {
+    for (auto it = shape.begin(); it != shape.end(); ++it) {
+      EXPECT_GT(*it, 0);
+    }
+    shape_ = std::vector<int32_t>(shape.begin(), shape.end());
+    size_ = LeakyReluTester::ComputeSize(shape_);
+    return *this;
+  }
+
+  inline const std::vector<int32_t>& Shape() const { return shape_; }
+
+  inline int32_t Size() const { return size_; }
+
+  inline LeakyReluTester& NegativeSlope(float negative_slope) {
+    negative_slope_ = negative_slope;
+    return *this;
+  }
+
+  inline float NegativeSlope() const { return negative_slope_; }
+
+  void Test(TfLiteDelegate* delegate) const;
+
+ private:
+  std::vector<char> CreateTfLiteModel() const;
+
+  static int32_t ComputeSize(const std::vector<int32_t>& shape);
+
+  std::vector<int32_t> shape_;
+  int32_t size_;
+  float negative_slope_ = 0.5f;
+};
+
+}  // namespace xnnpack
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_DELEGATES_XNNPACK_LEAKY_RELU_TESTER_H_

tensorflow/lite/delegates/xnnpack/xnnpack_delegate.cc

@@ -824,6 +824,14 @@ class Subgraph {
       case kTfLiteBuiltinHardSwish:
         return VisitHardSwishNode(subgraph, logging_context, node_index, node,
                                   context->tensors, xnnpack_tensors);
+      case kTfLiteBuiltinLeakyRelu: {
+        const TfLiteLeakyReluParams* leaky_relu_params =
+            static_cast<const TfLiteLeakyReluParams*>(node->builtin_data);
+
+        return VisitLeakyReluNode(subgraph, logging_context, node_index, node,
+                                  context->tensors, leaky_relu_params,
+                                  xnnpack_tensors);
+      }
       case kTfLiteBuiltinLogistic:
         return VisitLogisticNode(subgraph, logging_context, node_index, node,
                                  context->tensors, xnnpack_tensors);
@@ -1547,6 +1555,42 @@ class Subgraph {
     return kTfLiteOk;
   }
 
+  static TfLiteStatus VisitLeakyReluNode(
+      xnn_subgraph_t subgraph, TfLiteContext* logging_context, int node_index,
+      TfLiteNode* node, const TfLiteTensor* tensors,
+      const TfLiteLeakyReluParams* leaky_relu_params,
+      const std::vector<uint32_t>& xnnpack_tensors) {
+    TF_LITE_ENSURE_STATUS(
+        CheckNumInputsAndOutputs(logging_context, node, 1, 1, node_index));
+
+    const TfLiteTensor& input_tensor = tensors[node->inputs->data[0]];
+    TF_LITE_ENSURE_STATUS(CheckTensorFloatType(
+        logging_context, input_tensor, node->inputs->data[0], node_index));
+    TF_LITE_ENSURE_STATUS(CheckTensorNonDynamicAllocation(
+        logging_context, input_tensor, node->inputs->data[0], node_index));
+
+    const TfLiteTensor& output_tensor = tensors[node->outputs->data[0]];
+    TF_LITE_ENSURE_STATUS(CheckTensorFloatType(
+        logging_context, output_tensor, node->outputs->data[0], node_index));
+    TF_LITE_ENSURE_STATUS(CheckTensorNonDynamicAllocation(
+        logging_context, output_tensor, node->outputs->data[0], node_index));
+
+    if (subgraph != nullptr) {
+      const xnn_status status = xnn_define_leaky_relu(
+          subgraph, leaky_relu_params->alpha,
+          /*input_id=*/xnnpack_tensors[node->inputs->data[0]],
+          /*output_id=*/xnnpack_tensors[node->outputs->data[0]], /*flags=*/0);
+      if (status != xnn_status_success) {
+        TF_LITE_KERNEL_LOG(logging_context,
+                           "failed to delegate LEAKY_RELU node #%d",
+                           node_index);
+        return kTfLiteError;
+      }
+    }
+
+    return kTfLiteOk;
+  }
+
   static TfLiteStatus VisitLogisticNode(
       xnn_subgraph_t subgraph, TfLiteContext* logging_context, int node_index,
       TfLiteNode* node, const TfLiteTensor* tensors,
@@ -1572,7 +1616,7 @@ class Subgraph {
           /*output_id=*/xnnpack_tensors[node->outputs->data[0]], /*flags=*/0);
       if (status != xnn_status_success) {
         TF_LITE_KERNEL_LOG(logging_context,
-                           "failed to delegate SIGMOID node #%d", node_index);
+                           "failed to delegate LOGISTIC node #%d", node_index);
         return kTfLiteError;
       }
     }

tensorflow/workspace.bzl

@@ -164,11 +164,11 @@ def tf_repositories(path_prefix = "", tf_repo_name = ""):
 
     tf_http_archive(
         name = "XNNPACK",
-        sha256 = "ce44ddcf8b54d8ee87d7aee3a93a8cc949440547c41a8ab23fbf8a66314ef7ce",
-        strip_prefix = "XNNPACK-24567894452324bbc369e67b23fd86d87506511a",
+        sha256 = "3864557a77c84e94c4dad37c1b26cb0c232f3f7c6bf3210292f7e870b8b862df",
+        strip_prefix = "XNNPACK-5bbebac71014c5c2bafb65686f409bcf117baeb3",
         urls = [
-            "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/24567894452324bbc369e67b23fd86d87506511a.zip",
-            "https://github.com/google/XNNPACK/archive/24567894452324bbc369e67b23fd86d87506511a.zip",
+            "https://storage.googleapis.com/mirror.tensorflow.org/github.com/google/XNNPACK/archive/5bbebac71014c5c2bafb65686f409bcf117baeb3.zip",
+            "https://github.com/google/XNNPACK/archive/5bbebac71014c5c2bafb65686f409bcf117baeb3.zip",
         ],
     )