Rolling 20220103

code/components/tflite-lib/tensorflow/lite/kernels/internal/compatibility.h

@@ -0,0 +1,112 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_COMPATIBILITY_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_
+
+#include <cstdint>
+
+#include "tensorflow/lite/kernels/op_macros.h"
+
+#ifndef TFLITE_DCHECK
+#define TFLITE_DCHECK(condition) (condition) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_EQ
+#define TFLITE_DCHECK_EQ(x, y) ((x) == (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_NE
+#define TFLITE_DCHECK_NE(x, y) ((x) != (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_GE
+#define TFLITE_DCHECK_GE(x, y) ((x) >= (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_GT
+#define TFLITE_DCHECK_GT(x, y) ((x) > (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_LE
+#define TFLITE_DCHECK_LE(x, y) ((x) <= (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_LT
+#define TFLITE_DCHECK_LT(x, y) ((x) < (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+// TODO(ahentz): Clean up: We should stick to the DCHECK versions.
+#ifndef TFLITE_CHECK
+#define TFLITE_CHECK(condition) (condition) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_EQ
+#define TFLITE_CHECK_EQ(x, y) ((x) == (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_NE
+#define TFLITE_CHECK_NE(x, y) ((x) != (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_GE
+#define TFLITE_CHECK_GE(x, y) ((x) >= (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_GT
+#define TFLITE_CHECK_GT(x, y) ((x) > (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_LE
+#define TFLITE_CHECK_LE(x, y) ((x) <= (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_LT
+#define TFLITE_CHECK_LT(x, y) ((x) < (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TF_LITE_STATIC_MEMORY
+// TODO(b/162019032): Consider removing these type-aliases.
+using int8 = std::int8_t;
+using uint8 = std::uint8_t;
+using int16 = std::int16_t;
+using uint16 = std::uint16_t;
+using int32 = std::int32_t;
+using uint32 = std::uint32_t;
+#endif  // !defined(TF_LITE_STATIC_MEMORY)
+
+// TFLITE_DEPRECATED()
+//
+// Duplicated from absl/base/macros.h to avoid pulling in that library.
+// Marks a deprecated class, struct, enum, function, method and variable
+// declarations. The macro argument is used as a custom diagnostic message (e.g.
+// suggestion of a better alternative).
+//
+// Example:
+//
+//   class TFLITE_DEPRECATED("Use Bar instead") Foo {...};
+//   TFLITE_DEPRECATED("Use Baz instead") void Bar() {...}
+//
+// Every usage of a deprecated entity will trigger a warning when compiled with
+// clang's `-Wdeprecated-declarations` option. This option is turned off by
+// default, but the warnings will be reported by clang-tidy.
+#if defined(__clang__) && __cplusplus >= 201103L
+#define TFLITE_DEPRECATED(message) __attribute__((deprecated(message)))
+#endif
+
+#ifndef TFLITE_DEPRECATED
+#define TFLITE_DEPRECATED(message)
+#endif
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/cppmath.h

@@ -0,0 +1,40 @@
+/* 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_KERNELS_INTERNAL_CPPMATH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_
+
+#include <cmath>
+
+namespace tflite {
+
+#if defined(TF_LITE_USE_GLOBAL_CMATH_FUNCTIONS) || \
+    (defined(__ANDROID__) && !defined(__NDK_MAJOR__)) || defined(__ZEPHYR__)
+#define TF_LITE_GLOBAL_STD_PREFIX
+#else
+#define TF_LITE_GLOBAL_STD_PREFIX std
+#endif
+
+#define DECLARE_STD_GLOBAL_SWITCH1(tf_name, std_name) \
+  template <class T>                                  \
+  inline T tf_name(const T x) {                       \
+    return TF_LITE_GLOBAL_STD_PREFIX::std_name(x);    \
+  }
+
+DECLARE_STD_GLOBAL_SWITCH1(TfLiteRound, round);
+DECLARE_STD_GLOBAL_SWITCH1(TfLiteExpm1, expm1);
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/max.h

@@ -0,0 +1,35 @@
+/* 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_KERNELS_INTERNAL_MAX_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_
+
+#include <cmath>
+
+namespace tflite {
+
+#if defined(TF_LITE_USE_GLOBAL_MAX) || defined(__ZEPHYR__)
+inline float TfLiteMax(const float& x, const float& y) {
+  return std::max(x, y);
+}
+#else
+template <class T>
+inline T TfLiteMax(const T& x, const T& y) {
+  return std::fmax(x, y);
+}
+#endif
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/min.h

@@ -0,0 +1,35 @@
+/* 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_KERNELS_INTERNAL_MIN_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_
+
+#include <cmath>
+
+namespace tflite {
+
+#if defined(TF_LITE_USE_GLOBAL_MIN) || defined(__ZEPHYR__)
+inline float TfLiteMin(const float& x, const float& y) {
+  return std::min(x, y);
+}
+#else
+template <class T>
+inline T TfLiteMin(const T& x, const T& y) {
+  return std::fmin(x, y);
+}
+#endif
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/optimized/neon_check.h

@@ -0,0 +1,20 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_
+
+// TFLM does not need to utilize any Neon optimizations.
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/portable_tensor.h

@@ -0,0 +1,122 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_PORTABLE_TENSOR_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_
+
+#include <vector>
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+inline RuntimeShape GetTensorShape(std::vector<int32_t> data) {
+  return RuntimeShape(data.size(), data.data());
+}
+
+// A list of tensors in a format that can be used by kernels like split and
+// concatenation.
+template <typename T>
+class VectorOfTensors {
+ public:
+  // Build with the tensors in 'tensor_list'.
+  VectorOfTensors(const TfLiteContext& context,
+                  const TfLiteIntArray& tensor_list) {
+    int num_tensors = tensor_list.size;
+
+    all_data_.reserve(num_tensors);
+    all_shape_.reserve(num_tensors);
+    all_shape_ptr_.reserve(num_tensors);
+
+    for (int i = 0; i < num_tensors; ++i) {
+      TfLiteTensor* t = &context.tensors[tensor_list.data[i]];
+      all_data_.push_back(GetTensorData<T>(t));
+      all_shape_.push_back(GetTensorShape(t));
+    }
+
+    // Taking the pointer from inside a std::vector is only OK if the vector is
+    // never modified, so we populate all_shape in the previous loop and then we
+    // are free to grab iterators here.
+    for (int i = 0; i < num_tensors; ++i) {
+      all_shape_ptr_.push_back(&all_shape_[i]);
+    }
+  }
+  // Return a pointer to the data pointers of all tensors in the list. For
+  // example:
+  //   float* const* f = v.data();
+  //   f[0][1] is the second element of the first tensor.
+  T* const* data() const { return all_data_.data(); }
+
+  // Return a pointer the shape pointers of all tensors in the list. For
+  // example:
+  //   const RuntimeShape* const* d = v.dims();
+  //   dims[1] are the dimensions of the second tensor in the list.
+  const RuntimeShape* const* shapes() const { return all_shape_ptr_.data(); }
+
+ private:
+  std::vector<T*> all_data_;
+  std::vector<RuntimeShape> all_shape_;
+  std::vector<RuntimeShape*> all_shape_ptr_;
+};
+
+// A list of quantized tensors in a format that can be used by kernels like
+// split and concatenation.
+class VectorOfQuantizedTensors : public VectorOfTensors<uint8_t> {
+ public:
+  // Build with the tensors in 'tensor_list'.
+  VectorOfQuantizedTensors(const TfLiteContext& context,
+                           const TfLiteIntArray& tensor_list)
+      : VectorOfTensors<uint8_t>(context, tensor_list) {
+    for (int i = 0; i < tensor_list.size; ++i) {
+      TfLiteTensor* t = &context.tensors[tensor_list.data[i]];
+      zero_point_.push_back(t->params.zero_point);
+      scale_.push_back(t->params.scale);
+    }
+  }
+
+  const float* scale() const { return scale_.data(); }
+  const int32_t* zero_point() const { return zero_point_.data(); }
+
+ private:
+  std::vector<int32_t> zero_point_;
+  std::vector<float> scale_;
+};
+
+// Writes randomly accessed values from `input` sequentially into `output`.
+template <typename T>
+class SequentialTensorWriter {
+ public:
+  SequentialTensorWriter(const TfLiteTensor* input, TfLiteTensor* output) {
+    input_data_ = GetTensorData<T>(input);
+    output_ptr_ = GetTensorData<T>(output);
+  }
+  SequentialTensorWriter(const T* input_data, T* output_data)
+      : input_data_(input_data), output_ptr_(output_data) {}
+
+  void Write(int position) { *output_ptr_++ = input_data_[position]; }
+  void WriteN(int position, int len) {
+    memcpy(output_ptr_, &input_data_[position], sizeof(T) * len);
+    output_ptr_ += len;
+  }
+
+ private:
+  const T* input_data_;
+  T* output_ptr_;
+};
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/portable_tensor_utils.h

@@ -0,0 +1,124 @@
+/* Copyright 2021 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_KERNELS_INTERNAL_PORTABLE_TENSOR_UTILS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_UTILS_H_
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+
+#include "tensorflow/lite/c/builtin_op_data.h"
+#include "tensorflow/lite/c/common.h"
+
+#if defined(_MSC_VER)
+#define __restrict__ __restrict
+#endif
+
+namespace tflite {
+
+namespace tensor_utils {
+
+// Multiplies a matrix with a scalar and reduce the result on each row to a
+// scalar.
+// Parameters:
+//     - matrix: matrix of size n_row * n_col
+//     - scalar: the scalar that is multiplied to each element in the matrix
+//     - n_row:  the row count of the matrix
+//     - n_col:  the column count of the matrix
+//     - output: the 32bit output
+// Note: We do not need saturation because the int8 * int8 is safe from overflow
+// in (2^31-1) / (2^14) = 131072, which is bigger than the n_row. Non-zero
+// initial output value is not exceptionally large.
+void MatrixScalarMultiplyAccumulate(const int8_t* matrix, int32_t scalar,
+                                    int32_t n_row, int32_t n_col,
+                                    int32_t* output);
+
+// Add another vector for each batch in the batch vector.
+template <typename T>
+void VectorBatchVectorAdd(const T* vector, int v_size, int n_batch,
+                          T* batch_vector) {
+  for (int b = 0; b < n_batch; b++) {
+    for (int i = 0; i < v_size; ++i) {
+      batch_vector[i] += vector[i];
+    }
+    batch_vector += v_size;
+  }
+}
+
+// Cwise product of two vectors.
+template <typename T>
+inline void VectorVectorCwiseProduct(const T* __restrict__ vector1,
+                                     const T* __restrict__ vector2, int v_size,
+                                     T* __restrict__ result) {
+  for (int v = 0; v < v_size; v++) {
+    *result++ = *vector1++ * *vector2++;
+  }
+}
+
+// Cwise product of a vector and a batch-vector.
+template <typename T>
+inline void VectorBatchVectorCwiseProduct(const T* vector, int v_size,
+                                          const T* batch_vector, int n_batch,
+                                          T* result) {
+  for (int b = 0; b < n_batch; b++) {
+    VectorVectorCwiseProduct(vector, batch_vector, v_size, result);
+    // Update the pointers.
+    result += v_size;
+    batch_vector += v_size;
+  }
+}
+
+// Cwise product and accumulate of two vectors. Since it's a MAC operation, the
+// assumption here is that result array is initialized to valid values.
+template <typename T>
+inline void VectorVectorCwiseProductAccumulate(const T* __restrict__ vector1,
+                                               const T* __restrict__ vector2,
+                                               int v_size,
+                                               T* __restrict__ result) {
+  for (int v = 0; v < v_size; v++) {
+    *result++ += *vector1++ * *vector2++;
+  }
+}
+
+// Cwise product and accumulate of a vector and a batch-vector. Since it's a MAC
+// operation, the assumption here is that result array is initialized to valid
+// values.
+template <typename T>
+inline void VectorBatchVectorCwiseProductAccumulate(const T* vector, int v_size,
+                                                    const T* batch_vector,
+                                                    int n_batch, T* result) {
+  for (int b = 0; b < n_batch; b++) {
+    VectorVectorCwiseProductAccumulate(vector, batch_vector, v_size, result);
+    // Update the pointers.
+    result += v_size;
+    batch_vector += v_size;
+  }
+}
+
+// Batch vector initialization with another vector.
+template <typename T>
+void VectorBatchVectorAssign(const T* vector, int v_size, int n_batch,
+                             T* batch_vector) {
+  for (int b = 0; b < n_batch; b++) {
+    std::copy_n(vector, v_size, batch_vector + b * v_size);
+  }
+}
+
+}  // namespace tensor_utils
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_UTILS_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/quantization_util.cc

@@ -0,0 +1,416 @@
+/* Copyright 2017 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/kernels/internal/quantization_util.h"
+
+#include <algorithm>
+#include <cmath>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+
+namespace tflite {
+
+namespace {
+// These constants are used to manipulate the binary representation of doubles.
+// Double-precision binary64 floating point format is:
+// Bit |  63  |  62-52   |   51-0   |
+//     | Sign | Exponent | Fraction |
+// To avoid 64-bit integers as much as possible, I break this into high and
+// low 32-bit chunks. High is:
+// Bit |  31  |  30-20   |      19-0     |
+//     | Sign | Exponent | High Fraction |
+// Low is:
+// Bit |     31-0     |
+//     | Low Fraction |
+// We then access the components through logical bit-wise operations to
+// extract the parts needed, with the positions and masks derived from the
+// layout shown above.
+constexpr uint64_t kSignMask = 0x8000000000000000LL;
+constexpr uint64_t kExponentMask = 0x7ff0000000000000LL;
+constexpr int32_t kExponentShift = 52;
+constexpr int32_t kExponentBias = 1023;
+constexpr uint32_t kExponentIsBadNum = 0x7ff;
+constexpr uint64_t kFractionMask = 0x000fffffffc00000LL;
+constexpr uint32_t kFractionShift = 22;
+constexpr uint32_t kFractionRoundingMask = 0x003fffff;
+constexpr uint32_t kFractionRoundingThreshold = 0x00200000;
+}  // namespace
+
+void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier,
+                        int* shift) {
+#if TFLITE_SINGLE_ROUNDING
+  // Single-rounding MultiplyByQuantizedMultiplier only supports positive
+  // multipliers.
+  // TFLITE_DCHECK(double_multiplier >= 0);
+#endif
+  if (double_multiplier == 0.) {
+    *quantized_multiplier = 0;
+    *shift = 0;
+    return;
+  }
+#ifdef TFLITE_EMULATE_FLOAT
+  // If we're trying to avoid the use of floating-point instructions (for
+  // example on microcontrollers) then use an alternative implementation
+  // that only requires integer and bitwise operations. To enable this, you
+  // need to set the define during the build process for your platform.
+  int64_t q_fixed = IntegerFrExp(double_multiplier, shift);
+#else   // TFLITE_EMULATE_FLOAT
+  const double q = std::frexp(double_multiplier, shift);
+  auto q_fixed = static_cast<int64_t>(TfLiteRound(q * (1LL << 31)));
+#endif  // TFLITE_EMULATE_FLOAT
+  TFLITE_CHECK(q_fixed <= (1LL << 31));
+  if (q_fixed == (1LL << 31)) {
+    q_fixed /= 2;
+    ++*shift;
+  }
+  TFLITE_CHECK_LE(q_fixed, std::numeric_limits<int32_t>::max());
+  // A shift amount smaller than -31 would cause all bits to be shifted out
+  // and thus all results would be zero. We implement that instead with
+  // q_fixed==0, so as to avoid hitting issues with right-shift
+  // operations with shift amounts greater than 31. Note that this happens
+  // roughly when abs(double_multiplier) < 2^-31 and the present handling means
+  // that we're effectively flushing tiny double_multiplier's to zero.
+  // We could conceivably handle values in the range (roughly) [32, 63]
+  // as 'denormals' i.e. (shift==0, q_fixed < 2^30). In that point of view
+  // the present handling is just doing 'flush denormals to zero'. We could
+  // reconsider and actually generate nonzero denormals if a need arises.
+  if (*shift < -31) {
+    *shift = 0;
+    q_fixed = 0;
+  }
+#if TFLITE_SINGLE_ROUNDING
+  // Single-rounding MultiplyByQuantizedMultiplier doesn't support a shift > 30,
+  // saturate it.
+  if (*shift > 30) {
+    *shift = 30;
+    q_fixed = (1LL << 31) - 1;
+  }
+#endif
+  *quantized_multiplier = static_cast<int32_t>(q_fixed);
+}
+
+void QuantizeMultiplierGreaterThanOne(double double_multiplier,
+                                      int32_t* quantized_multiplier,
+                                      int* left_shift) {
+  TFLITE_CHECK_GT(double_multiplier, 1.);
+  QuantizeMultiplier(double_multiplier, quantized_multiplier, left_shift);
+  TFLITE_CHECK_GE(*left_shift, 0);
+}
+
+void QuantizeMultiplierSmallerThanOneExp(double double_multiplier,
+                                         int32_t* quantized_multiplier,
+                                         int* left_shift) {
+  TFLITE_CHECK_LT(double_multiplier, 1.);
+  TFLITE_CHECK_GT(double_multiplier, 0.);
+  int shift;
+  QuantizeMultiplier(double_multiplier, quantized_multiplier, &shift);
+  TFLITE_CHECK_LE(shift, 0);
+  *left_shift = shift;
+}
+
+int64_t IntegerFrExp(double input, int* shift) {
+  // Make sure our assumptions about the double layout hold.
+  TFLITE_CHECK_EQ(8, sizeof(double));
+
+  // We want to access the bits of the input double value directly, which is
+  // tricky to do safely, so use a union to handle the casting.
+  union {
+    double double_value;
+    uint64_t double_as_uint;
+  } cast_union;
+  cast_union.double_value = input;
+  const uint64_t u = cast_union.double_as_uint;
+
+  // If the bitfield is all zeros apart from the sign bit, this is a normalized
+  // zero value, so return standard values for this special case.
+  if ((u & ~kSignMask) == 0) {
+    *shift = 0;
+    return 0;
+  }
+
+  // Deal with NaNs and Infs, which are always indicated with a fixed pattern in
+  // the exponent, and distinguished by whether the fractions are zero or
+  // non-zero.
+  const uint32_t exponent_part = ((u & kExponentMask) >> kExponentShift);
+  if (exponent_part == kExponentIsBadNum) {
+    *shift = std::numeric_limits<int>::max();
+    if (u & kFractionMask) {
+      // NaN, so just return zero (with the exponent set to INT_MAX).
+      return 0;
+    } else {
+      // Infinity, so return +/- INT_MAX.
+      if (u & kSignMask) {
+        return std::numeric_limits<int64_t>::min();
+      } else {
+        return std::numeric_limits<int64_t>::max();
+      }
+    }
+  }
+
+  // The shift is fairly easy to extract from the high bits of the double value,
+  // just by masking it out and applying a bias. The std::frexp() implementation
+  // always returns values between 0.5 and 1.0 though, whereas the exponent
+  // assumes 1.0 to 2.0 is the standard range, so I add on one to match that
+  // interface.
+  *shift = (exponent_part - kExponentBias) + 1;
+
+  // There's an implicit high bit in the double format definition, so make sure
+  // we include that at the top, and then reconstruct the rest of the fractional
+  // value from the remaining fragments.
+  int64_t fraction = 0x40000000 + ((u & kFractionMask) >> kFractionShift);
+
+  // We're cutting off some bits at the bottom, so to exactly match the standard
+  // frexp implementation here we'll apply rounding by adding one to the least
+  // significant bit of the result if the discarded portion is over half of the
+  // maximum.
+  if ((u & kFractionRoundingMask) > kFractionRoundingThreshold) {
+    fraction += 1;
+  }
+  // Negate the fraction if the sign bit was set.
+  if (u & kSignMask) {
+    fraction *= -1;
+  }
+
+  return fraction;
+}
+
+double DoubleFromFractionAndShift(int64_t fraction, int shift) {
+  union {
+    double double_value;
+    uint64_t double_as_uint;
+  } result;
+
+  // Detect NaNs and infinities.
+  if (shift == std::numeric_limits<int>::max()) {
+    if (fraction == 0) {
+      return std::numeric_limits<double>::quiet_NaN();
+    } else if (fraction > 0) {
+      return std::numeric_limits<double>::infinity();
+    } else {
+      return -std::numeric_limits<double>::infinity();
+    }
+  }
+
+  // Return a normalized zero for a zero fraction.
+  if (fraction == 0) {
+    result.double_as_uint = 0;
+    return result.double_value;
+  }
+
+  bool is_negative = (fraction < 0);
+  int64_t encoded_fraction = is_negative ? -fraction : fraction;
+  int64_t encoded_shift = (shift - 1);
+  while (encoded_fraction < 0x40000000) {
+    encoded_fraction *= 2;
+    encoded_shift -= 1;
+  }
+  while (encoded_fraction > 0x80000000) {
+    encoded_fraction /= 2;
+    encoded_shift += 1;
+  }
+  encoded_fraction -= 0x40000000;
+  if (encoded_shift < -1022) {
+    encoded_shift = -1023;
+  } else if (encoded_shift > 1022) {
+    encoded_shift = 1023;
+  }
+  encoded_shift += kExponentBias;
+  uint64_t encoded_sign = is_negative ? kSignMask : 0;
+  result.double_as_uint = encoded_sign | (encoded_shift << kExponentShift) |
+                          (encoded_fraction << kFractionShift);
+  return result.double_value;
+}
+
+double IntegerDoubleMultiply(double a, double b) {
+  int a_shift;
+  const int64_t a_fraction = IntegerFrExp(a, &a_shift);
+  int b_shift;
+  const int64_t b_fraction = IntegerFrExp(b, &b_shift);
+  // Detect NaNs and infinities.
+  if (a_shift == std::numeric_limits<int>::max() ||
+      (b_shift == std::numeric_limits<int>::max())) {
+    return std::numeric_limits<double>::quiet_NaN();
+  }
+  const int result_shift = a_shift + b_shift + 1;
+  const int64_t result_fraction = (a_fraction * b_fraction) >> 32;
+  return DoubleFromFractionAndShift(result_fraction, result_shift);
+}
+
+int IntegerDoubleCompare(double a, double b) {
+  int a_shift;
+  const int64_t a_fraction = IntegerFrExp(a, &a_shift);
+  int b_shift;
+  const int64_t b_fraction = IntegerFrExp(b, &b_shift);
+
+  // Detect NaNs and infinities.
+  if (a_shift == std::numeric_limits<int>::max() ||
+      (b_shift == std::numeric_limits<int>::max())) {
+    return 1;
+  }
+
+  if ((a_fraction == 0) && (b_fraction < 0)) {
+    return 1;
+  } else if ((a_fraction < 0) && (b_fraction == 0)) {
+    return -1;
+  } else if (a_shift < b_shift) {
+    return -1;
+  } else if (a_shift > b_shift) {
+    return 1;
+  } else if (a_fraction < b_fraction) {
+    return -1;
+  } else if (a_fraction > b_fraction) {
+    return 1;
+  } else {
+    return 0;
+  }
+}
+
+void PreprocessSoftmaxScaling(double beta, double input_scale,
+                              int input_integer_bits,
+                              int32_t* quantized_multiplier, int* left_shift) {
+  // If the overall multiplier (input and beta) is large, then exp() of an
+  // input difference of 1 scaled by this will be large.  In other words, we
+  // can cap the multiplier and know that, when it is used, the output will be
+  // (round to) zero wherever the input is not at the maximum value.
+
+  // If the overall scale is less than one, and input_integer_bits=0, then the
+  // result is double equivalent of Q0.31 (actually with more precision). Thus
+  // this generates a Q(input_integer_bits).(31-input_integer_bits)
+  // representation.
+#if TFLITE_SINGLE_ROUNDING
+  const double max_real_multiplier = (1LL << 30) - 1.0;
+#else
+  const double max_real_multiplier = (1LL << 31) - 1.0;
+#endif
+
+#ifdef TFLITE_EMULATE_FLOAT
+  const double input_beta = IntegerDoubleMultiply(beta, input_scale);
+  int shift;
+  int64_t fraction = IntegerFrExp(input_beta, &shift);
+  shift += (31 - input_integer_bits);
+  double input_beta_real_multiplier =
+      DoubleFromFractionAndShift(fraction, shift);
+  if (IntegerDoubleCompare(input_beta_real_multiplier, max_real_multiplier) >
+      0) {
+    input_beta_real_multiplier = max_real_multiplier;
+  }
+#else   // TFLITE_EMULATE_FLOAT
+  const double input_beta_real_multiplier =
+      std::min<double>(beta * input_scale * (1 << (31 - input_integer_bits)),
+                       max_real_multiplier);
+#endif  // TFLITE_EMULATE_FLOAT
+
+  QuantizeMultiplierGreaterThanOne(input_beta_real_multiplier,
+                                   quantized_multiplier, left_shift);
+}
+
+void PreprocessLogSoftmaxScalingExp(double beta, double input_scale,
+                                    int input_integer_bits,
+                                    int32_t* quantized_multiplier,
+                                    int* left_shift,
+                                    int32_t* reverse_scaling_divisor,
+                                    int* reverse_scaling_left_shift) {
+  PreprocessSoftmaxScaling(beta, input_scale, input_integer_bits,
+                           quantized_multiplier, left_shift);
+
+  // Also calculate what amounts to the inverse scaling factor for the input.
+  const double real_reverse_scaling_divisor =
+      (1 << (31 - *left_shift)) / static_cast<double>(*quantized_multiplier);
+  tflite::QuantizeMultiplierSmallerThanOneExp(real_reverse_scaling_divisor,
+                                              reverse_scaling_divisor,
+                                              reverse_scaling_left_shift);
+}
+
+int CalculateInputRadius(int input_integer_bits, int input_left_shift,
+                         int total_signed_bits) {
+#ifdef TFLITE_EMULATE_FLOAT
+  int64_t result = (1 << input_integer_bits) - 1;
+  result <<= (total_signed_bits - input_integer_bits);
+  result >>= input_left_shift;
+  return result;
+#else   // TFLITE_EMULATE_FLOAT
+  const double max_input_rescaled =
+      1.0 * ((1 << input_integer_bits) - 1) *
+      (1LL << (total_signed_bits - input_integer_bits)) /
+      (1LL << input_left_shift);
+  // Tighten bound using floor.  Suppose that we could use the exact value.
+  // After scaling the difference, the result would be at the maximum.  Thus we
+  // must ensure that our value has lower magnitude.
+  return static_cast<int>(std::floor(max_input_rescaled));
+#endif  // TFLITE_EMULATE_FLOAT
+}
+
+void NudgeQuantizationRange(const float min, const float max,
+                            const int quant_min, const int quant_max,
+                            float* nudged_min, float* nudged_max,
+                            float* nudged_scale) {
+  // This code originates from tensorflow/core/kernels/fake_quant_ops_functor.h.
+  const float quant_min_float = static_cast<float>(quant_min);
+  const float quant_max_float = static_cast<float>(quant_max);
+  *nudged_scale = (max - min) / (quant_max_float - quant_min_float);
+  const float zero_point_from_min = quant_min_float - min / *nudged_scale;
+  uint16_t nudged_zero_point;
+  if (zero_point_from_min < quant_min_float) {
+    nudged_zero_point = static_cast<uint16_t>(quant_min);
+  } else if (zero_point_from_min > quant_max_float) {
+    nudged_zero_point = static_cast<uint16_t>(quant_max);
+  } else {
+    nudged_zero_point = static_cast<uint16_t>(TfLiteRound(zero_point_from_min));
+  }
+  *nudged_min = (quant_min_float - nudged_zero_point) * (*nudged_scale);
+  *nudged_max = (quant_max_float - nudged_zero_point) * (*nudged_scale);
+}
+
+void FakeQuantizeArray(const float nudged_scale, const float nudged_min,
+                       const float nudged_max, const float* input_data,
+                       float* output_data, const float size) {
+  // This code originates from tensorflow/core/kernels/fake_quant_ops_functor.h.
+  const float inv_nudged_scale = 1.0f / nudged_scale;
+
+  for (int i = 0; i < size; i++) {
+    const float src_val = input_data[i];
+    const float clamped = std::min(nudged_max, std::max(nudged_min, src_val));
+    const float clamped_shifted = clamped - nudged_min;
+    const float dst_val =
+        TfLiteRound(clamped_shifted * inv_nudged_scale) * nudged_scale +
+        nudged_min;
+    output_data[i] = dst_val;
+  }
+}
+
+bool CheckedLog2(const float x, int* log2_result) {
+  // Using TfLiteRound instead of std::round and std::log instead of
+  // std::log2 to work around these functions being missing in a toolchain
+  // used in some TensorFlow tests as of May 2018.
+  const float x_log2 = std::log(x) * (1.0f / std::log(2.0f));
+  const float x_log2_rounded = TfLiteRound(x_log2);
+  const float x_log2_fracpart = x_log2 - x_log2_rounded;
+
+  *log2_result = static_cast<int>(x_log2_rounded);
+  return std::abs(x_log2_fracpart) < 1e-3f;
+}
+
+void QuantizeMultiplierArray(const double* effective_scales, size_t size,
+                             int32_t* effective_scale_significand,
+                             int* effective_shift) {
+  for (size_t i = 0; i < size; ++i) {
+    QuantizeMultiplier(effective_scales[i], &effective_scale_significand[i],
+                       &effective_shift[i]);
+  }
+}
+
+}  // namespace tflite

code/components/tflite-lib/tensorflow/lite/kernels/internal/quantization_util.h

@@ -0,0 +1,292 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_
+
+#include <cmath>
+#include <cstdint>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+// Given the min and max values of a float array, return
+// reasonable quantization parameters to use for this array.
+template <typename T>
+QuantizationParams ChooseQuantizationParams(double rmin, double rmax,
+                                            bool narrow_range) {
+  const T qmin = std::numeric_limits<T>::min() + (narrow_range ? 1 : 0);
+  const T qmax = std::numeric_limits<T>::max();
+  const double qmin_double = qmin;
+  const double qmax_double = qmax;
+  // 0 should always be a representable value. Let's assume that the initial
+  // min,max range contains 0.
+  TFLITE_CHECK_LE(rmin, 0.);
+  TFLITE_CHECK_GE(rmax, 0.);
+  if (rmin == rmax) {
+    // Special case where the min,max range is a point. Should be {0}.
+    TFLITE_CHECK_EQ(rmin, 0.);
+    TFLITE_CHECK_EQ(rmax, 0.);
+    QuantizationParams quantization_params;
+    quantization_params.zero_point = 0;
+    quantization_params.scale = 0.;
+    return quantization_params;
+  }
+
+  // General case.
+  //
+  // First determine the scale.
+  const double scale = (rmax - rmin) / (qmax_double - qmin_double);
+
+  // Zero-point computation.
+  // First the initial floating-point computation. The zero-point can be
+  // determined from solving an affine equation for any known pair
+  // (real value, corresponding quantized value).
+  // We know two such pairs: (rmin, qmin) and (rmax, qmax).
+  // The arithmetic error on the zero point computed from either pair
+  // will be roughly machine_epsilon * (sum of absolute values of terms)
+  // so we want to use the variant that adds the smaller terms.
+  const double zero_point_from_min = qmin_double - rmin / scale;
+  const double zero_point_from_max = qmax_double - rmax / scale;
+  const double zero_point_from_min_error =
+      std::abs(qmin_double) + std::abs(rmin / scale);
+  const double zero_point_from_max_error =
+      std::abs(qmax_double) + std::abs(rmax / scale);
+
+  const double zero_point_double =
+      zero_point_from_min_error < zero_point_from_max_error
+          ? zero_point_from_min
+          : zero_point_from_max;
+
+  // Now we need to nudge the zero point to be an integer
+  // (our zero points are integer, and this is motivated by the requirement
+  // to be able to represent the real value "0" exactly as a quantized value,
+  // which is required in multiple places, for example in Im2col with SAME
+  // padding).
+  T nudged_zero_point = 0;
+  if (zero_point_double < qmin_double) {
+    nudged_zero_point = qmin;
+  } else if (zero_point_double > qmax_double) {
+    nudged_zero_point = qmax;
+  } else {
+    nudged_zero_point = static_cast<T>(round(zero_point_double));
+  }
+  // The zero point should always be in the range of quantized value,
+  // [qmin, qmax].
+  TFLITE_CHECK_GE(nudged_zero_point, qmin);
+  TFLITE_CHECK_LE(nudged_zero_point, qmax);
+
+  // Finally, store the result nudged quantization params.
+  QuantizationParams quantization_params;
+  quantization_params.zero_point = nudged_zero_point;
+  quantization_params.scale = scale;
+  return quantization_params;
+}
+
+template <typename T>
+QuantizationParams ChooseQuantizationParams(double rmin, double rmax) {
+  return ChooseQuantizationParams<T>(rmin, rmax, false);
+}
+
+// Converts a floating-point number to an integer. For all inputs x where
+// static_cast<IntOut>(x) is legal according to the C++ standard, the result
+// is identical to that cast (i.e. the result is x with its fractional part
+// truncated whenever that is representable as IntOut).
+//
+// static_cast would cause undefined behavior for the following cases, which
+// have well-defined behavior for this function:
+//
+//  1. If x is NaN, the result is zero.
+//
+//  2. If the truncated form of x is above the representable range of IntOut,
+//     the result is std::numeric_limits<IntOut>::max().
+//
+//  3. If the truncated form of x is below the representable range of IntOut,
+//     the result is std::numeric_limits<IntOut>::min().
+//
+// Note that cases #2 and #3 cover infinities as well as finite numbers.
+//
+// The range of FloatIn must include the range of IntOut, otherwise
+// the results are undefined.
+// TODO(sfeuz): Replace by absl::SafeCast once available.
+template <class IntOut, class FloatIn>
+IntOut SafeCast(FloatIn x) {
+  static_assert(!std::numeric_limits<FloatIn>::is_integer,
+                "FloatIn is integer");
+  static_assert(std::numeric_limits<IntOut>::is_integer,
+                "IntOut is not integer");
+  static_assert(std::numeric_limits<IntOut>::radix == 2, "IntOut is base 2");
+
+  // Special case NaN, for which the logic below doesn't work.
+  if (std::isnan(x)) {
+    return 0;
+  }
+
+  // Negative values all clip to zero for unsigned results.
+  if (!std::numeric_limits<IntOut>::is_signed && x < 0) {
+    return 0;
+  }
+
+  // Handle infinities.
+  if (std::isinf(x)) {
+    return x < 0 ? std::numeric_limits<IntOut>::min()
+                 : std::numeric_limits<IntOut>::max();
+  }
+
+  // Set exp such that x == f * 2^exp for some f with |f| in [0.5, 1.0),
+  // unless x is zero in which case exp == 0. Note that this implies that the
+  // magnitude of x is strictly less than 2^exp.
+  int exp = 0;
+  std::frexp(x, &exp);
+
+  // Let N be the number of non-sign bits in the representation of IntOut. If
+  // the magnitude of x is strictly less than 2^N, the truncated version of x
+  // is representable as IntOut. The only representable integer for which this
+  // is not the case is kMin for signed types (i.e. -2^N), but that is covered
+  // by the fall-through below.
+  if (exp <= std::numeric_limits<IntOut>::digits) {
+    return x;
+  }
+
+  // Handle numbers with magnitude >= 2^N.
+  return x < 0 ? std::numeric_limits<IntOut>::min()
+               : std::numeric_limits<IntOut>::max();
+}
+
+// Decompose a double multiplier into a Q0.31 int32 representation of its
+// significand, and shift representation of NEGATIVE its exponent ---
+// this is intended as a RIGHT-shift.
+//
+// Restricted to the case where the multiplier < 1 (and non-negative).
+void QuantizeMultiplierSmallerThanOneExp(double double_multiplier,
+                                         int32_t* quantized_multiplier,
+                                         int* left_shift);
+
+// Decompose a double multiplier into a Q0.31 int32 representation of its
+// significand, and shift representation of its exponent.
+//
+// Restricted to the case where the multiplier > 1.
+void QuantizeMultiplierGreaterThanOne(double double_multiplier,
+                                      int32_t* quantized_multiplier,
+                                      int* left_shift);
+
+// Decompose a double multiplier into a Q0.31 int32 representation of its
+// significand, and shift representation of its exponent.
+//
+// Handles an arbitrary positive multiplier. The 'shift' output-value is
+// basically the 'floating-point exponent' of the multiplier:
+// Negative for a right-shift (when the multiplier is <1), positive for a
+// left-shift (when the multiplier is >1)
+void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier,
+                        int* shift);
+
+// Splits a double input value into a returned fraction, and a shift value from
+// the exponent, using only bitwise and integer operations to support
+// microcontrollers and other environments without floating-point support.
+//
+// This is designed to be a replacement for how std::frexp() is used within the
+// QuantizeMultiplier() function, and so has a different signature than the
+// standard version, returning a 64-bit integer rather than a double. This
+// result has a maximum value of 1<<31, with the fraction expressed as a
+// proportion of that maximum.
+//
+// std::frexp() returns NaNs and infinities unmodified, but since we're
+// returning integers that can't represent those values, instead we return
+// a shift of std::numeric_limits<int>::max() for all bad numbers, with an int64
+// result of 0 for NaNs, std:numeric_limits<int64_t>::max() for +INFINITY, and
+// std::numeric_limits<int64_t>::min() for -INFINITY. Denormalized inputs will
+// result in return values that end up truncating some bits at the end,
+// reflecting the loss of precision inherent in denormalization.
+int64_t IntegerFrExp(double input, int* shift);
+
+// Converts an integer fraction in the format produced by IntegerFrExp (where
+// 0x40000000 is 1.0) and an exponent shift (between -1022 and +1022) into an
+// IEEE binary64 double format result. The implementation uses only integer and
+// bitwise operators, so no floating point hardware support or emulation is
+// needed. This is here so quantized operations can run non-time-critical
+// preparation calculations on microcontrollers and other platforms without
+// float support.
+double DoubleFromFractionAndShift(int64_t fraction, int shift);
+
+// Performs a multiplication of two numbers in double format, using only integer
+// and bitwise instructions. This is aimed at supporting housekeeping functions
+// for quantized operations on microcontrollers without floating-point hardware.
+double IntegerDoubleMultiply(double a, double b);
+
+// Returns -1 if a is less than b, 0 if a and b are equal, and +1 if a is
+// greater than b. It is implemented using only integer and logical instructions
+// so that it can be easily run on microcontrollers for quantized operations.
+int IntegerDoubleCompare(double a, double b);
+
+// This first creates a multiplier in a double equivalent of
+// Q(input_integer_bits).(31-input_integer_bits) representation, with extra
+// precision in the double's fractional bits.  It then splits the result into
+// significand and exponent.
+void PreprocessSoftmaxScaling(double beta, double input_scale,
+                              int input_integer_bits,
+                              int32_t* quantized_multiplier, int* left_shift);
+// Like PreprocessSoftmaxScaling, but inverse scaling factors also calculated.
+void PreprocessLogSoftmaxScalingExp(double beta, double input_scale,
+                                    int input_integer_bits,
+                                    int32_t* quantized_multiplier,
+                                    int* left_shift,
+                                    int32_t* reverse_scaling_divisor,
+                                    int* reverse_scaling_left_shift);
+// Calculate the largest input that will result in a within-bounds intermediate
+// result within MultiplyByQuantizedMultiplierGreaterThanOne.  In other words,
+// it must not overflow before we reduce the value by multiplication by the
+// input multiplier.  The negative radius is used as the minimum difference in
+// Softmax.
+int CalculateInputRadius(int input_integer_bits, int input_left_shift,
+                         int total_signed_bits = 31);
+
+// Nudges a min/max quantization range to ensure zero is zero.
+// Gymnastics with nudged zero point is to ensure that real zero maps to
+// an integer, which is required for e.g. zero-padding in convolutional layers.
+// Outputs nudged_min, nudged_max, nudged_scale.
+void NudgeQuantizationRange(const float min, const float max,
+                            const int quant_min, const int quant_max,
+                            float* nudged_min, float* nudged_max,
+                            float* nudged_scale);
+
+// Fake quantizes (quantizes and dequantizes) input_data using the scale,
+// nudged_min, and nudged_max from NudgeQuantizationRange. This matches the code
+// in TensorFlow's FakeQuantizeWithMinMaxVarsFunctor.
+void FakeQuantizeArray(const float nudged_scale, const float nudged_min,
+                       const float nudged_max, const float* input_data,
+                       float* output_data, const float size);
+
+// If x is approximately a power of two (with any positive or negative
+// exponent), stores that exponent (i.e. log2(x)) in *log2_result, otherwise
+// returns false.
+bool CheckedLog2(const float x, int* log2_result);
+
+// Decomposes an array of double multipliers into a Q0.31 int32 representation
+// of its significand, and shift representation of its exponent.
+//
+// Handles an arbitrary multiplier. The 'shift' output-value is
+// basically the 'floating-point exponent' of the multiplier:
+// Negative for a right-shift (when the multiplier is <1), positive for a
+// left-shift (when the multiplier is >1)
+void QuantizeMultiplierArray(const double* effective_scales, size_t size,
+                             int32_t* effective_scale_significand,
+                             int* effective_shift);
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/add.h

@@ -0,0 +1,399 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_ADD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_
+
+#include <type_traits>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const T* input1_data,
+                const RuntimeShape& input2_shape, const T* input2_data,
+                const RuntimeShape& output_shape, T* output_data) {
+  T activation_min, activation_max;
+  GetActivationParams(params, &activation_min, &activation_max);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] + input2_data[i], activation_min, activation_max);
+  }
+}
+
+// Element-wise add that can often be used for inner loop of broadcast add as
+// well as the non-broadcast add.
+
+// This function is used for 8-bit as well as for 16-bit, but the accumulator
+// is 32-bit for both cases. The overflow does not happen due to the
+// choice of the shift (20 or 15, accordingly - see add.cc for more comments).
+template <typename T>
+inline void AddElementwise(int size, const ArithmeticParams& params,
+                           const T* input1_data, const T* input2_data,
+                           T* output_data) {
+  TFLITE_DCHECK_GT(params.input1_offset, -std::numeric_limits<T>::max());
+  TFLITE_DCHECK_GT(params.input2_offset, -std::numeric_limits<T>::max());
+  TFLITE_DCHECK_LT(params.input1_offset, std::numeric_limits<T>::max());
+  TFLITE_DCHECK_LT(params.input2_offset, std::numeric_limits<T>::max());
+
+  for (int i = 0; i < size; ++i) {
+    const int32_t input1_val = params.input1_offset + input1_data[i];
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+    const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+    const int32_t scaled_input1_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input1_val, params.input1_multiplier, params.input1_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, params.input2_multiplier, params.input2_shift);
+    const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+    const int32_t raw_output =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            raw_sum, params.output_multiplier, params.output_shift) +
+        params.output_offset;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+// Scalar-broadcast add that can be used for inner loop of more general
+// broadcast add, so that, for example, scalar-broadcast with batch will still
+// be fast.
+inline void AddScalarBroadcast(int size, const ArithmeticParams& params,
+                               uint8_t input1_data, const uint8_t* input2_data,
+                               uint8_t* output_data) {
+  TFLITE_DCHECK_GT(params.input1_offset, -256);
+  TFLITE_DCHECK_GT(params.input2_offset, -256);
+  TFLITE_DCHECK_LT(params.input1_offset, 256);
+  TFLITE_DCHECK_LT(params.input2_offset, 256);
+
+  const int32_t input1_val = params.input1_offset + input1_data;
+  const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+  const int32_t scaled_input1_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input1_val, params.input1_multiplier, params.input1_shift);
+  for (int i = 0; i < size; ++i) {
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, params.input2_multiplier, params.input2_shift);
+    const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+    const int32_t raw_output =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            raw_sum, params.output_multiplier, params.output_shift) +
+        params.output_offset;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[i] = static_cast<uint8_t>(clamped_output);
+  }
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const uint8_t* input1_data,
+                const RuntimeShape& input2_shape, const uint8_t* input2_data,
+                const RuntimeShape& output_shape, uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  TFLITE_DCHECK_GT(params.input1_offset, -256);
+  TFLITE_DCHECK_GT(params.input2_offset, -256);
+  TFLITE_DCHECK_LT(params.input1_offset, 256);
+  TFLITE_DCHECK_LT(params.input2_offset, 256);
+  AddElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void AddGeneralParamScale(const ArithmeticParams& params,
+                                 const RuntimeShape& input1_shape,
+                                 const int16_t* input1_data,
+                                 const RuntimeShape& input2_shape,
+                                 const int16_t* input2_data,
+                                 const RuntimeShape& output_shape,
+                                 int16_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  int max_value = std::numeric_limits<int16_t>::max();
+
+  TFLITE_DCHECK_GT(params.input1_offset, -max_value);
+  TFLITE_DCHECK_GT(params.input2_offset, -max_value);
+  TFLITE_DCHECK_LT(params.input1_offset, max_value);
+  TFLITE_DCHECK_LT(params.input2_offset, max_value);
+  AddElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int16_t* input1_data,
+                const RuntimeShape& input2_shape, const int16_t* input2_data,
+                const RuntimeShape& output_shape, int16_t* output_data,
+                bool pot_scale = true) {
+  if (!pot_scale) {
+    AddGeneralParamScale(params, input1_shape, input1_data, input2_shape,
+                         input2_data, output_shape, output_data);
+    return;
+  }
+
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+
+  const int input1_shift = params.input1_shift;
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  const int16_t output_activation_min = params.quantized_activation_min;
+  const int16_t output_activation_max = params.quantized_activation_max;
+
+  TFLITE_DCHECK(input1_shift == 0 || params.input2_shift == 0);
+  TFLITE_DCHECK_LE(input1_shift, 0);
+  TFLITE_DCHECK_LE(params.input2_shift, 0);
+  const int16_t* not_shift_input =
+      input1_shift == 0 ? input1_data : input2_data;
+  const int16_t* shift_input = input1_shift == 0 ? input2_data : input1_data;
+  const int input_right_shift =
+      input1_shift == 0 ? -params.input2_shift : -input1_shift;
+
+  for (int i = 0; i < flat_size; i++) {
+    // F0 uses 0 integer bits, range [-1, 1].
+    using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+
+    F0 input_ready_scaled = F0::FromRaw(not_shift_input[i]);
+    F0 scaled_input = F0::FromRaw(
+        gemmlowp::RoundingDivideByPOT(shift_input[i], input_right_shift));
+    F0 result = gemmlowp::SaturatingAdd(scaled_input, input_ready_scaled);
+    const int16_t raw_output = result.raw();
+    const int16_t clamped_output = std::min(
+        output_activation_max, std::max(output_activation_min, raw_output));
+    output_data[i] = clamped_output;
+  }
+}
+
+template <typename T>
+inline typename std::enable_if<!is_small_integer<T>::value, void>::type
+BroadcastAdd4DSlow(const ArithmeticParams& params,
+                   const RuntimeShape& input1_shape, const T* input1_data,
+                   const RuntimeShape& input2_shape, const T* input2_data,
+                   const RuntimeShape& output_shape, T* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  T activation_min, activation_max;
+  GetActivationParams(params, &activation_min, &activation_max);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              ActivationFunctionWithMinMax<T>(
+                  input1_data[SubscriptToIndex(desc1, b, y, x, c)] +
+                      input2_data[SubscriptToIndex(desc2, b, y, x, c)],
+                  activation_min, activation_max);
+        }
+      }
+    }
+  }
+}
+
+// This function is used for 8-bit as well as for 16-bit, but the accumulator
+// is 32-bit for both cases. The overflow does not happen due to the
+// choice of the shift (20 or 15, accordingly - see add.cc for more comments).
+template <typename T>
+inline typename std::enable_if<is_small_integer<T>::value, void>::type
+BroadcastAdd4DSlow(const ArithmeticParams& params,
+                   const RuntimeShape& input1_shape, const T* input1_data,
+                   const RuntimeShape& input2_shape, const T* input2_data,
+                   const RuntimeShape& output_shape, T* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              params.input1_offset +
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)];
+          const int32_t input2_val =
+              params.input2_offset +
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+          const int32_t shifted_input1_val =
+              input1_val * (1 << params.left_shift);
+          const int32_t shifted_input2_val =
+              input2_val * (1 << params.left_shift);
+          const int32_t scaled_input1_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input1_val, params.input1_multiplier,
+                  params.input1_shift);
+          const int32_t scaled_input2_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input2_val, params.input2_multiplier,
+                  params.input2_shift);
+          const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+          const int32_t raw_output =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  raw_sum, params.output_multiplier, params.output_shift) +
+              params.output_offset;
+          const int32_t clamped_output =
+              std::min(params.quantized_activation_max,
+                       std::max(params.quantized_activation_min, raw_output));
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              static_cast<T>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+inline void BroadcastAddFivefold(const ArithmeticParams& unswitched_params,
+                                 const RuntimeShape& unswitched_input1_shape,
+                                 const uint8_t* unswitched_input1_data,
+                                 const RuntimeShape& unswitched_input2_shape,
+                                 const uint8_t* unswitched_input2_data,
+                                 const RuntimeShape& output_shape,
+                                 uint8_t* output_data) {
+  ArithmeticParams switched_params = unswitched_params;
+  switched_params.input1_offset = unswitched_params.input2_offset;
+  switched_params.input1_multiplier = unswitched_params.input2_multiplier;
+  switched_params.input1_shift = unswitched_params.input2_shift;
+  switched_params.input2_offset = unswitched_params.input1_offset;
+  switched_params.input2_multiplier = unswitched_params.input1_multiplier;
+  switched_params.input2_shift = unswitched_params.input1_shift;
+
+  const bool use_unswitched =
+      unswitched_params.broadcast_category ==
+      tflite::BroadcastableOpCategory::kFirstInputBroadcastsFast;
+
+  const ArithmeticParams& params =
+      use_unswitched ? unswitched_params : switched_params;
+  const uint8_t* input1_data =
+      use_unswitched ? unswitched_input1_data : unswitched_input2_data;
+  const uint8_t* input2_data =
+      use_unswitched ? unswitched_input2_data : unswitched_input1_data;
+
+  // Fivefold nested loops. The second input resets its position for each
+  // iteration of the second loop. The first input resets its position at the
+  // beginning of the fourth loop. The innermost loop is an elementwise add of
+  // sections of the arrays.
+  uint8_t* output_data_ptr = output_data;
+  const uint8_t* input1_data_ptr = input1_data;
+  const uint8_t* input2_data_reset = input2_data;
+  // In the fivefold pattern, y0, y2 and y4 are not broadcast, and so shared
+  // between input shapes. y3 for input 1 is always broadcast, and so the
+  // dimension there is 1, whereas optionally y1 might be broadcast for input 2.
+  // Put another way,
+  // input1.shape.FlatSize = y0 * y1 * y2 * y4,
+  // input2.shape.FlatSize = y0 * y2 * y3 * y4.
+  int y0 = params.broadcast_shape[0];
+  int y1 = params.broadcast_shape[1];
+  int y2 = params.broadcast_shape[2];
+  int y3 = params.broadcast_shape[3];
+  int y4 = params.broadcast_shape[4];
+  if (y4 > 1) {
+    // General fivefold pattern, with y4 > 1 so there is a non-broadcast inner
+    // dimension.
+    for (int i0 = 0; i0 < y0; ++i0) {
+      const uint8_t* input2_data_ptr;
+      for (int i1 = 0; i1 < y1; ++i1) {
+        input2_data_ptr = input2_data_reset;
+        for (int i2 = 0; i2 < y2; ++i2) {
+          for (int i3 = 0; i3 < y3; ++i3) {
+            AddElementwise(y4, params, input1_data_ptr, input2_data_ptr,
+                           output_data_ptr);
+            input2_data_ptr += y4;
+            output_data_ptr += y4;
+          }
+          // We have broadcast y4 of input1 data y3 times, and now move on.
+          input1_data_ptr += y4;
+        }
+      }
+      // We have broadcast y2*y3*y4 of input2 data y1 times, and now move on.
+      input2_data_reset = input2_data_ptr;
+    }
+  } else {
+    // Special case of y4 == 1, in which the innermost loop is a single element
+    // and can be combined with the next (y3) as an inner broadcast.
+    //
+    // Note that this handles the case of pure scalar broadcast when
+    // y0 == y1 == y2 == 1. With low overhead it handles cases such as scalar
+    // broadcast with batch (as y2 > 1).
+    //
+    // NOTE The process is the same as the above general case except simplified
+    // for y4 == 1 and the loop over y3 is contained within the
+    // AddScalarBroadcast function.
+    for (int i0 = 0; i0 < y0; ++i0) {
+      const uint8_t* input2_data_ptr;
+      for (int i1 = 0; i1 < y1; ++i1) {
+        input2_data_ptr = input2_data_reset;
+        for (int i2 = 0; i2 < y2; ++i2) {
+          AddScalarBroadcast(y3, params, *input1_data_ptr, input2_data_ptr,
+                             output_data_ptr);
+          input2_data_ptr += y3;
+          output_data_ptr += y3;
+          input1_data_ptr += 1;
+        }
+      }
+      input2_data_reset = input2_data_ptr;
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/add_n.h

@@ -0,0 +1,86 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_ADD_N_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_N_H_
+
+#include <algorithm>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_ops {
+
+// T is expected to be either float or int.
+template <typename T>
+inline void AddN(const RuntimeShape& input_shape, const size_t num_inputs,
+                 const T* const* input_data, T* output_data) {
+  // All inputs and output should have the same shape, this is checked during
+  // Prepare stage.
+  const size_t size = input_shape.FlatSize();
+  for (size_t i = 0; i < size; ++i) {
+    T x = 0;
+    for (size_t j = 0; j < num_inputs; ++j) {
+      x += input_data[j][i];
+    }
+    output_data[i] = x;
+  }
+}
+
+inline void AddN(const ArithmeticParams& params,
+                 const RuntimeShape& input_shape, const size_t num_inputs,
+                 const int8_t* const* input_data, int8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  // Input offset is negative input zero point. Activation tensors are
+  // asymmetric quantized so they span the full int8 range.
+  // All inputs should have same zero-point and scale, this is checked during
+  // Prepare stage.
+  TFLITE_DCHECK_GE(-params.input1_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_LE(-params.input1_offset, std::numeric_limits<int8_t>::max());
+
+  // All inputs and output should have the same shape, this is checked during
+  // Prepare stage.
+  const size_t size = input_shape.FlatSize();
+  for (size_t i = 0; i < size; ++i) {
+    // accumulate in scaled_x before clamping to avoid overflow
+    const int32_t x = params.input1_offset;  // x = 0
+    const int32_t shifted_x = x * (1 << params.left_shift);
+    int32_t scaled_x = MultiplyByQuantizedMultiplierSmallerThanOneExp(
+        shifted_x, params.input1_multiplier, params.input1_shift);
+
+    for (size_t j = 0; j < num_inputs; ++j) {
+      const int32_t y = params.input1_offset + input_data[j][i];
+      const int32_t shifted_y = y * (1 << params.left_shift);
+      int32_t scaled_y = MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_y, params.input1_multiplier, params.input1_shift);
+      scaled_x += scaled_y;
+    }
+
+    const int32_t raw_output =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            scaled_x, params.output_multiplier, params.output_shift) +
+        params.output_offset;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[i] = static_cast<int8_t>(clamped_output);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_N_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/arg_min_max.h

@@ -0,0 +1,88 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_
+
+#include <functional>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+std::function<bool(T, T)> GetComparefunction(bool is_arg_max) {
+  if (is_arg_max) {
+    return std::greater<T>();
+  } else {
+    return std::less<T>();
+  }
+}
+
+template <typename T1, typename T2, typename T3, typename Cmp>
+void ArgMinMax(const RuntimeShape& input1_shape, const T1* input1_data,
+               const T3* input2_data, const RuntimeShape& output_shape,
+               T2* output_data, const Cmp& cmp) {
+  TFLITE_DCHECK_GT(input1_shape.DimensionsCount(), 0);
+  TFLITE_DCHECK_EQ(input1_shape.DimensionsCount() - 1,
+                   output_shape.DimensionsCount());
+  int axis = input2_data[0];
+  if (axis < 0) {
+    axis += input1_shape.DimensionsCount();
+  }
+  const int axis_size = input1_shape.Dims(axis);
+
+  int outer_size = 1;
+  for (int i = 0; i < axis; ++i) {
+    TFLITE_DCHECK_EQ(input1_shape.Dims(i), output_shape.Dims(i));
+    outer_size *= input1_shape.Dims(i);
+  }
+
+  int inner_size = 1;
+  const int dims_count = input1_shape.DimensionsCount();
+  for (int i = axis + 1; i < dims_count; ++i) {
+    TFLITE_DCHECK_EQ(input1_shape.Dims(i), output_shape.Dims(i - 1));
+    inner_size *= input1_shape.Dims(i);
+  }
+  for (int outer = 0; outer < outer_size; ++outer) {
+    for (int inner = 0; inner < inner_size; ++inner) {
+      auto min_max_value = input1_data[outer * axis_size * inner_size + inner];
+      T2 min_max_index = 0;
+      for (int i = 1; i < axis_size; ++i) {
+        const auto& curr_value =
+            input1_data[(outer * axis_size + i) * inner_size + inner];
+        if (cmp(curr_value, min_max_value)) {
+          min_max_value = curr_value;
+          min_max_index = static_cast<T2>(i);
+        }
+      }
+      output_data[outer * inner_size + inner] = min_max_index;
+    }
+  }
+}
+
+template <typename T1, typename T2, typename T3>
+void ArgMinMax(const RuntimeShape& input1_shape, const T1* input1_data,
+               const T3* input2_data, const RuntimeShape& output_shape,
+               T2* output_data, const bool is_arg_max) {
+  ArgMinMax(input1_shape, input1_data, input2_data, output_shape, output_data,
+            GetComparefunction<T1>(is_arg_max));
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/batch_matmul.h

@@ -0,0 +1,275 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_BATCH_MATMUL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BATCH_MATMUL_H_
+
+#include <algorithm>
+#include <cstdint>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/tensor_utils_common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+namespace batch_matmul {
+
+// Determine which dimension is the broadcast dimension.
+inline int broadcast_dim(int lhs_dim, int rhs_dim) {
+  if (lhs_dim == rhs_dim) return lhs_dim;
+  if (lhs_dim == 1) return rhs_dim;
+  TFLITE_DCHECK_EQ(rhs_dim, 1);
+  return lhs_dim;
+}
+
+// Compute the "extent" for iterating on this dimension.
+// If we are broadcasting, then don't advance (i.e return 0).
+inline int extent(const RuntimeShape& shape, int x) {
+  if (shape.Dims(x) == 1) {
+    return 0;
+  }
+  int prod = 1;
+  for (int i = x + 1; i < shape.DimensionsCount(); ++i) {
+    prod *= shape.Dims(i);
+  }
+  return prod;
+}
+
+}  // namespace batch_matmul
+
+template <typename Ta, typename Tb, typename Tout>
+inline void BatchMatMul(const RuntimeShape& lhs_shape, const Ta* lhs_data,
+                        const RuntimeShape& rhs_shape, const Tb* rhs_data,
+                        const RuntimeShape& output_shape, Tout* output_data) {
+  const RuntimeShape extended_lhs_shape =
+      RuntimeShape::ExtendedShape(5, lhs_shape);
+  const RuntimeShape extended_rhs_shape =
+      RuntimeShape::ExtendedShape(5, rhs_shape);
+
+  const int batch_dim0 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
+  const int batch_dim1 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
+  const int batch_dim2 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));
+
+  const int lhs_ext0 = batch_matmul::extent(extended_lhs_shape, 0);
+  const int lhs_ext1 = batch_matmul::extent(extended_lhs_shape, 1);
+  const int lhs_ext2 = batch_matmul::extent(extended_lhs_shape, 2);
+  const int rhs_ext0 = batch_matmul::extent(extended_rhs_shape, 0);
+  const int rhs_ext1 = batch_matmul::extent(extended_rhs_shape, 1);
+  const int rhs_ext2 = batch_matmul::extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const int lhs_rows = extended_lhs_shape.Dims(3);
+  const int rhs_cols = extended_rhs_shape.Dims(4);
+  const int accum_depth = extended_lhs_shape.Dims(4);
+
+  for (int b0 = 0; b0 < batch_dim0; ++b0) {
+    const Ta* lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
+    const Tb* rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
+    for (int b1 = 0; b1 < batch_dim1; ++b1) {
+      const Ta* lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
+      const Tb* rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
+      for (int b2 = 0; b2 < batch_dim2; ++b2) {
+        const Ta* lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
+        const Tb* rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
+        Tout* out_ptr = output_data + ((b0 * batch_dim1 * batch_dim2) +
+                                       b1 * batch_dim2 + b2) *
+                                          lhs_rows * rhs_cols;
+        for (int j = 0; j < rhs_cols; ++j) {
+          for (int i = 0; i < lhs_rows; ++i) {
+            Tout total = 0;
+            for (int k = 0; k < accum_depth; ++k) {
+              total += static_cast<Tout>(lhs_ptr2[accum_depth * i + k]) *
+                       static_cast<Tout>(rhs_ptr2[j * accum_depth + k]);
+            }
+            int idx = lhs_rows * j + i;
+            out_ptr[idx] = total;
+          }
+        }
+      }
+    }
+  }
+}
+
+inline void BatchMatMul(const RuntimeShape& lhs_shape, const int8_t* lhs_data,
+                        const RuntimeShape& rhs_shape, const int8_t* rhs_data,
+                        const float* scaling_factors,
+                        const int32_t* input_offset, int32_t* row_sums,
+                        const RuntimeShape& output_shape, float* output_data,
+                        bool* compute_row_sums) {
+  const RuntimeShape extended_lhs_shape =
+      RuntimeShape::ExtendedShape(5, lhs_shape);
+  const RuntimeShape extended_rhs_shape =
+      RuntimeShape::ExtendedShape(5, rhs_shape);
+
+  const int batch_dim0 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
+  const int batch_dim1 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
+  const int batch_dim2 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));
+
+  const int lhs_ext0 = batch_matmul::extent(extended_lhs_shape, 0);
+  const int lhs_ext1 = batch_matmul::extent(extended_lhs_shape, 1);
+  const int lhs_ext2 = batch_matmul::extent(extended_lhs_shape, 2);
+  const int rhs_ext0 = batch_matmul::extent(extended_rhs_shape, 0);
+  const int rhs_ext1 = batch_matmul::extent(extended_rhs_shape, 1);
+  const int rhs_ext2 = batch_matmul::extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const int lhs_rows = extended_lhs_shape.Dims(3);
+  const int rhs_cols = extended_rhs_shape.Dims(4);
+  const int accum_depth = extended_lhs_shape.Dims(4);
+
+  const int ioff_ext0 = rhs_ext0 == 0 ? 0 : rhs_cols;
+  const int ioff_ext1 = rhs_ext1 == 0 ? 0 : rhs_cols;
+  const int ioff_ext2 = rhs_ext2 == 0 ? 0 : rhs_cols;
+  const int woff_ext0 = lhs_ext0 == 0 ? 0 : lhs_rows;
+  const int woff_ext1 = lhs_ext1 == 0 ? 0 : lhs_rows;
+  const int woff_ext2 = lhs_ext2 == 0 ? 0 : lhs_rows;
+
+  if (!compute_row_sums || *compute_row_sums) {
+    int num_weights_matrices = 1;
+    for (int i = 1; i < extended_lhs_shape.DimensionsCount() - 2; ++i) {
+      num_weights_matrices *= extended_lhs_shape.Dims(i);
+    }
+    tensor_utils::ReductionSumVector(
+        lhs_data, row_sums, num_weights_matrices * lhs_rows, accum_depth);
+    if (compute_row_sums) {
+      *compute_row_sums = false;
+    }
+  }
+
+  for (int b0 = 0; b0 < batch_dim0; ++b0) {
+    const int8_t* lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
+    const int8_t* rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
+    const int32_t* ioff_ptr0 = input_offset + (b0 * ioff_ext0);
+    const float* scale_ptr0 = scaling_factors + (b0 * ioff_ext0);
+    const int32_t* woff_ptr0 = row_sums + (b0 * woff_ext0);
+    for (int b1 = 0; b1 < batch_dim1; ++b1) {
+      const int8_t* lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
+      const int8_t* rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
+      const int32_t* ioff_ptr1 = ioff_ptr0 + (b1 * ioff_ext1);
+      const float* scale_ptr1 = scale_ptr0 + (b1 * ioff_ext1);
+      const int32_t* woff_ptr1 = woff_ptr0 + (b1 * woff_ext1);
+      for (int b2 = 0; b2 < batch_dim2; ++b2) {
+        const int8_t* lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
+        const int8_t* rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
+        const int32_t* ioff_ptr2 = ioff_ptr1 + (b2 * ioff_ext2);
+        const float* scale_ptr2 = scale_ptr1 + (b2 * ioff_ext2);
+        const int32_t* woff_ptr2 = woff_ptr1 + (b2 * woff_ext2);
+        float* out_ptr = output_data + ((b0 * batch_dim1 * batch_dim2) +
+                                        b1 * batch_dim2 + b2) *
+                                           lhs_rows * rhs_cols;
+        for (int j = 0; j < rhs_cols; ++j) {
+          const float batch_scaling_factor = scale_ptr2[j];
+          const float batch_offset = static_cast<float>(ioff_ptr2[j]);
+          for (int i = 0; i < lhs_rows; ++i) {
+            int32_t total = 0;
+            for (int k = 0; k < accum_depth; ++k) {
+              total +=
+                  lhs_ptr2[accum_depth * i + k] * rhs_ptr2[j * accum_depth + k];
+            }
+            int32_t row_sum = woff_ptr2[i];
+            total -= row_sum * batch_offset;
+            int idx = lhs_rows * j + i;
+            out_ptr[idx] += batch_scaling_factor * total;
+          }
+        }
+      }
+    }
+  }
+}
+
+template <typename T, typename AccumT>
+inline void BatchMatMul(const FullyConnectedParams& params,
+                        const RuntimeShape& lhs_shape, const T* lhs_data,
+                        const RuntimeShape& rhs_shape, const T* rhs_data,
+                        const RuntimeShape& output_shape, T* output_data) {
+  const RuntimeShape extended_lhs_shape =
+      RuntimeShape::ExtendedShape(5, lhs_shape);
+  const RuntimeShape extended_rhs_shape =
+      RuntimeShape::ExtendedShape(5, rhs_shape);
+
+  const int batch_dim0 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0));
+  const int batch_dim1 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1));
+  const int batch_dim2 = batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2));
+
+  const int lhs_ext0 = batch_matmul::extent(extended_lhs_shape, 0);
+  const int lhs_ext1 = batch_matmul::extent(extended_lhs_shape, 1);
+  const int lhs_ext2 = batch_matmul::extent(extended_lhs_shape, 2);
+  const int rhs_ext0 = batch_matmul::extent(extended_rhs_shape, 0);
+  const int rhs_ext1 = batch_matmul::extent(extended_rhs_shape, 1);
+  const int rhs_ext2 = batch_matmul::extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const int lhs_rows = extended_lhs_shape.Dims(3);
+  const int rhs_cols = extended_rhs_shape.Dims(4);
+  const int accum_depth = extended_lhs_shape.Dims(4);
+
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  for (int b0 = 0; b0 < batch_dim0; ++b0) {
+    const T* lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
+    const T* rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
+    for (int b1 = 0; b1 < batch_dim1; ++b1) {
+      const T* lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
+      const T* rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
+      for (int b2 = 0; b2 < batch_dim2; ++b2) {
+        const T* lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
+        const T* rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
+        T* out_ptr = output_data +
+                     ((b0 * batch_dim1 * batch_dim2) + b1 * batch_dim2 + b2) *
+                         lhs_rows * rhs_cols;
+
+        for (int j = 0; j < rhs_cols; ++j) {
+          for (int i = 0; i < lhs_rows; ++i) {
+            AccumT total = 0;
+            for (int k = 0; k < accum_depth; ++k) {
+              AccumT lhs_val = lhs_ptr2[accum_depth * i + k];
+              AccumT rhs_val = rhs_ptr2[accum_depth * j + k];
+              total += (lhs_val + filter_offset) * (rhs_val + input_offset);
+            }
+            int32_t total_scaled = MultiplyByQuantizedMultiplier(
+                total, output_multiplier, output_shift);
+            total_scaled += output_offset;
+            total_scaled = std::max(total_scaled, output_activation_min);
+            total_scaled = std::min(total_scaled, output_activation_max);
+            const int idx = lhs_rows * j + i;
+            out_ptr[idx] = static_cast<T>(total_scaled);
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BATCH_MATMUL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/batch_to_space_nd.h

@@ -0,0 +1,101 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_BATCH_TO_SPACE_ND_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BATCH_TO_SPACE_ND_H_
+
+#include <cmath>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+// TODO(b/135760455): Move this method anonymous namespace in a cc file.
+inline RuntimeShape ExtendShapeBatchToSpace(const RuntimeShape& shape) {
+  if (shape.DimensionsCount() == 4) {
+    return shape;
+  }
+  RuntimeShape new_shape(4, 1);
+  new_shape.SetDim(0, shape.Dims(0));
+  new_shape.SetDim(1, shape.Dims(1));
+  new_shape.SetDim(3, shape.Dims(2));
+  return new_shape;
+}
+
+template <typename T>
+inline void BatchToSpaceND(const RuntimeShape& unextended_input1_shape,
+                           const T* input1_data,
+                           const RuntimeShape& unextended_input2_shape,
+                           const int32_t* block_shape_data,
+                           const RuntimeShape& unextended_input3_shape,
+                           const int32_t* crops_data,
+                           const RuntimeShape& unextended_output_shape,
+                           T* output_data) {
+  ruy::profiler::ScopeLabel label("BatchToSpaceND");
+  TFLITE_DCHECK_GE(unextended_input1_shape.DimensionsCount(), 3);
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(unextended_input1_shape.DimensionsCount(),
+                   unextended_output_shape.DimensionsCount());
+
+  const RuntimeShape input1_shape =
+      ExtendShapeBatchToSpace(unextended_input1_shape);
+  const RuntimeShape output_shape =
+      ExtendShapeBatchToSpace(unextended_output_shape);
+
+  const int output_width = output_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_batch_size = output_shape.Dims(0);
+
+  const int depth = input1_shape.Dims(3);
+  const int input_width = input1_shape.Dims(2);
+  const int input_height = input1_shape.Dims(1);
+  const int input_batch_size = input1_shape.Dims(0);
+
+  const int block_shape_height = block_shape_data[0];
+  const int block_shape_width =
+      unextended_input1_shape.DimensionsCount() == 4 ? block_shape_data[1] : 1;
+  const int crops_top = crops_data[0];
+  const int crops_left =
+      unextended_input1_shape.DimensionsCount() == 4 ? crops_data[2] : 0;
+  for (int in_batch = 0; in_batch < input_batch_size; ++in_batch) {
+    const int out_batch = in_batch % output_batch_size;
+    const int spatial_offset = in_batch / output_batch_size;
+    for (int in_h = 0; in_h < input_height; ++in_h) {
+      const int out_h = in_h * block_shape_height +
+                        spatial_offset / block_shape_width - crops_top;
+      if (out_h < 0 || out_h >= output_height) {
+        continue;
+      }
+      for (int in_w = 0; in_w < input_width; ++in_w) {
+        const int out_w = in_w * block_shape_width +
+                          spatial_offset % block_shape_width - crops_left;
+
+        if (out_w < 0 || out_w >= output_width) {
+          continue;
+        }
+        T* out = output_data + Offset(output_shape, out_batch, out_h, out_w, 0);
+        const T* in =
+            input1_data + Offset(input1_shape, in_batch, in_h, in_w, 0);
+        memcpy(out, in, depth * sizeof(T));
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BATCH_TO_SPACE_ND_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/binary_function.h

@@ -0,0 +1,91 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Also appears to duplicate MinimumMaximum.
+//
+// R: Result type. T1: Input 1 type. T2: Input 2 type.
+template <typename R, typename T1, typename T2>
+inline void BroadcastBinaryFunction4DSlow(
+    const RuntimeShape& unextended_input1_shape, const T1* input1_data,
+    const RuntimeShape& unextended_input2_shape, const T2* input2_data,
+    const RuntimeShape& unextended_output_shape, R* output_data,
+    R (*func)(T1, T2)) {
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
+                                      unextended_input2_shape, &desc1, &desc2);
+
+  const int* dims_data =
+      reinterpret_cast<const int*>(output_shape.DimsDataUpTo5D());
+  for (int b = 0; b < output_shape.Dims(0); ++b) {
+    int out_idx_b = b * dims_data[1];
+    int in_idx1_b = desc1.strides[0] * b;
+    int in_idx2_b = desc2.strides[0] * b;
+    for (int y = 0; y < output_shape.Dims(1); ++y) {
+      int out_idx_y = (out_idx_b + y) * dims_data[2];
+      int in_idx1_y = in_idx1_b + desc1.strides[1] * y;
+      int in_idx2_y = in_idx2_b + desc2.strides[1] * y;
+      for (int x = 0; x < output_shape.Dims(2); ++x) {
+        int out_idx_x = (out_idx_y + x) * dims_data[3];
+        int in1_idx = in_idx1_y + desc1.strides[2] * x;
+        int in2_idx = in_idx2_y + desc2.strides[2] * x;
+        for (int c = 0; c < output_shape.Dims(3); ++c) {
+          auto out_idx = out_idx_x + c;
+          auto in1_val = input1_data[in1_idx];
+          auto in2_val = input2_data[in2_idx];
+          output_data[out_idx] = func(in1_val, in2_val);
+          in1_idx += desc1.strides[3];
+          in2_idx += desc2.strides[3];
+        }
+      }
+    }
+  }
+}
+
+// R: Result type. T1: Input 1 type. T2: Input 2 type.
+template <typename R, typename T1, typename T2>
+inline void BinaryFunction(const RuntimeShape& input1_shape,
+                           const T1* input1_data,
+                           const RuntimeShape& input2_shape,
+                           const T2* input2_data,
+                           const RuntimeShape& output_shape, R* output_data,
+                           R (*func)(T1, T2)) {
+  const int flat_size =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = func(input1_data[i], input2_data[i]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/ceil.h

@@ -0,0 +1,37 @@
+/* Copyright 2018 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_KERNELS_INTERNAL_REFERENCE_CEIL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void Ceil(const RuntimeShape& input_shape, const float* input_data,
+                 const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = std::ceil(input_data[i]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/comparisons.h

@@ -0,0 +1,280 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline bool EqualFn(T lhs, T rhs) {
+  return lhs == rhs;
+}
+
+template <typename T>
+inline bool NotEqualFn(T lhs, T rhs) {
+  return lhs != rhs;
+}
+
+template <typename T>
+inline bool GreaterFn(T lhs, T rhs) {
+  return lhs > rhs;
+}
+template <typename T>
+inline bool GreaterEqualFn(T lhs, T rhs) {
+  return lhs >= rhs;
+}
+template <typename T>
+inline bool LessFn(T lhs, T rhs) {
+  return lhs < rhs;
+}
+template <typename T>
+inline bool LessEqualFn(T lhs, T rhs) {
+  return lhs <= rhs;
+}
+
+template <typename T>
+using ComparisonFn = bool (*)(T, T);
+
+template <typename T, ComparisonFn<T> F>
+inline void ComparisonImpl(
+    const ComparisonParams& op_params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, bool* output_data) {
+  const int64_t flatsize =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+  for (int64_t i = 0; i < flatsize; ++i) {
+    output_data[i] = F(input1_data[i], input2_data[i]);
+  }
+}
+
+template <ComparisonFn<float> F>
+inline void Comparison(const ComparisonParams& op_params,
+                       const RuntimeShape& input1_shape,
+                       const float* input1_data,
+                       const RuntimeShape& input2_shape,
+                       const float* input2_data,
+                       const RuntimeShape& output_shape, bool* output_data) {
+  ComparisonImpl<float, F>(op_params, input1_shape, input1_data, input2_shape,
+                           input2_data, output_shape, output_data);
+}
+
+template <typename T, ComparisonFn<int32_t> F>
+inline void ComparisonWithScaling(
+    const ComparisonParams& op_params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, bool* output_data) {
+  int left_shift = op_params.left_shift;
+  int32_t input1_offset = op_params.input1_offset;
+  int32_t input1_multiplier = op_params.input1_multiplier;
+  int input1_shift = op_params.input1_shift;
+  int32_t input2_offset = op_params.input2_offset;
+  int32_t input2_multiplier = op_params.input2_multiplier;
+  int input2_shift = op_params.input2_shift;
+
+  const int64_t flatsize =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+  for (int64_t i = 0; i < flatsize; ++i) {
+    const int32_t input1_val = input1_offset + input1_data[i];
+    const int32_t input2_val = input2_offset + input2_data[i];
+    const int32_t shifted_input1_val = input1_val * (1 << left_shift);
+    const int32_t shifted_input2_val = input2_val * (1 << left_shift);
+    const int32_t scaled_input1_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input1_val, input1_multiplier, input1_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, input2_multiplier, input2_shift);
+    output_data[i] = F(scaled_input1_val, scaled_input2_val);
+  }
+}
+
+struct BroadcastComparison4DSlowCommon {
+  const RuntimeShape output_shape;
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+};
+
+inline BroadcastComparison4DSlowCommon BroadcastComparison4DSlowPreprocess(
+    const RuntimeShape& unextended_input1_shape,
+    const RuntimeShape& unextended_input2_shape,
+    const RuntimeShape& unextended_output_shape) {
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
+                                      unextended_input2_shape, &desc1, &desc2);
+  return {RuntimeShape::ExtendedShape(4, unextended_output_shape), desc1,
+          desc2};
+}
+
+template <typename T, ComparisonFn<T> F>
+inline void BroadcastComparison4DSlowImpl(
+    const ComparisonParams& op_params,
+    const RuntimeShape& unextended_input1_shape, const T* input1_data,
+    const RuntimeShape& unextended_input2_shape, const T* input2_data,
+    const RuntimeShape& unextended_output_shape, bool* output_data) {
+  const BroadcastComparison4DSlowCommon dims =
+      BroadcastComparison4DSlowPreprocess(unextended_input1_shape,
+                                          unextended_input2_shape,
+                                          unextended_output_shape);
+
+  for (int b = 0; b < dims.output_shape.Dims(0); ++b) {
+    for (int y = 0; y < dims.output_shape.Dims(1); ++y) {
+      for (int x = 0; x < dims.output_shape.Dims(2); ++x) {
+        for (int c = 0; c < dims.output_shape.Dims(3); ++c) {
+          output_data[Offset(dims.output_shape, b, y, x, c)] =
+              F(input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)],
+                input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)]);
+        }
+      }
+    }
+  }
+}
+
+template <ComparisonFn<float> F>
+inline void BroadcastComparison4DSlow(const ComparisonParams& op_params,
+                                      const RuntimeShape& input1_shape,
+                                      const float* input1_data,
+                                      const RuntimeShape& input2_shape,
+                                      const float* input2_data,
+                                      const RuntimeShape& output_shape,
+                                      bool* output_data) {
+  BroadcastComparison4DSlowImpl<float, F>(op_params, input1_shape, input1_data,
+                                          input2_shape, input2_data,
+                                          output_shape, output_data);
+}
+
+template <typename T, ComparisonFn<int32_t> F>
+inline void BroadcastComparison4DSlowWithScaling(
+    const ComparisonParams& op_params,
+    const RuntimeShape& unextended_input1_shape, const T* input1_data,
+    const RuntimeShape& unextended_input2_shape, const T* input2_data,
+    const RuntimeShape& unextended_output_shape, bool* output_data) {
+  const BroadcastComparison4DSlowCommon dims =
+      BroadcastComparison4DSlowPreprocess(unextended_input1_shape,
+                                          unextended_input2_shape,
+                                          unextended_output_shape);
+
+  int left_shift = op_params.left_shift;
+  int32_t input1_offset = op_params.input1_offset;
+  int32_t input1_multiplier = op_params.input1_multiplier;
+  int input1_shift = op_params.input1_shift;
+  int32_t input2_offset = op_params.input2_offset;
+  int32_t input2_multiplier = op_params.input2_multiplier;
+  int input2_shift = op_params.input2_shift;
+
+  for (int b = 0; b < dims.output_shape.Dims(0); ++b) {
+    for (int y = 0; y < dims.output_shape.Dims(1); ++y) {
+      for (int x = 0; x < dims.output_shape.Dims(2); ++x) {
+        for (int c = 0; c < dims.output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              input1_offset +
+              input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)];
+          const int32_t input2_val =
+              input2_offset +
+              input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)];
+          const int32_t shifted_input1_val = input1_val * (1 << left_shift);
+          const int32_t shifted_input2_val = input2_val * (1 << left_shift);
+          const int32_t scaled_input1_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input1_val, input1_multiplier, input1_shift);
+          const int32_t scaled_input2_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input2_val, input2_multiplier, input2_shift);
+          output_data[Offset(dims.output_shape, b, y, x, c)] =
+              F(scaled_input1_val, scaled_input2_val);
+        }
+      }
+    }
+  }
+}
+
+#define TFLITE_COMPARISON_OP(name)                                             \
+  inline void name(const ComparisonParams& op_params,                          \
+                   const RuntimeShape& input1_shape, const float* input1_data, \
+                   const RuntimeShape& input2_shape, const float* input2_data, \
+                   const RuntimeShape& output_shape, bool* output_data) {      \
+    Comparison<name##Fn>(op_params, input1_shape, input1_data, input2_shape,   \
+                         input2_data, output_shape, output_data);              \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void name##NoScaling(                                                 \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    ComparisonImpl<T, name##Fn>(op_params, input1_shape, input1_data,          \
+                                input2_shape, input2_data, output_shape,       \
+                                output_data);                                  \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void name##WithScaling(                                               \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    ComparisonWithScaling<T, name##Fn>(op_params, input1_shape, input1_data,   \
+                                       input2_shape, input2_data,              \
+                                       output_shape, output_data);             \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void Broadcast4DSlow##name##NoScaling(                                \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    BroadcastComparison4DSlowImpl<T, name##Fn>(                                \
+        op_params, input1_shape, input1_data, input2_shape, input2_data,       \
+        output_shape, output_data);                                            \
+  }                                                                            \
+  inline void Broadcast4DSlow##name(                                           \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const float* input1_data, const RuntimeShape& input2_shape,              \
+      const float* input2_data, const RuntimeShape& output_shape,              \
+      bool* output_data) {                                                     \
+    BroadcastComparison4DSlow<name##Fn>(op_params, input1_shape, input1_data,  \
+                                        input2_shape, input2_data,             \
+                                        output_shape, output_data);            \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void Broadcast4DSlow##name##WithScaling(                              \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    BroadcastComparison4DSlowWithScaling<T, name##Fn>(                         \
+        op_params, input1_shape, input1_data, input2_shape, input2_data,       \
+        output_shape, output_data);                                            \
+  }
+TFLITE_COMPARISON_OP(Equal);
+TFLITE_COMPARISON_OP(NotEqual);
+TFLITE_COMPARISON_OP(Greater);
+TFLITE_COMPARISON_OP(GreaterEqual);
+TFLITE_COMPARISON_OP(Less);
+TFLITE_COMPARISON_OP(LessEqual);
+#undef TFLITE_COMPARISON_OP
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/concatenation.h

@@ -0,0 +1,139 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename Scalar>
+inline void Concatenation(const ConcatenationParams& params,
+                          const RuntimeShape* const* input_shapes,
+                          const Scalar* const* input_data,
+                          const RuntimeShape& output_shape,
+                          Scalar* output_data) {
+  int axis = params.axis;
+  int inputs_count = params.inputs_count;
+  const int concat_dimensions = output_shape.DimensionsCount();
+  TFLITE_DCHECK_LT(axis, concat_dimensions);
+
+  int64_t concat_size = 0;
+  for (int i = 0; i < inputs_count; i++) {
+    TFLITE_DCHECK_EQ(input_shapes[i]->DimensionsCount(), concat_dimensions);
+    for (int j = 0; j < concat_dimensions; j++) {
+      if (j != axis) {
+        MatchingDim(*input_shapes[i], j, output_shape, j);
+      }
+    }
+    concat_size += input_shapes[i]->Dims(axis);
+  }
+  TFLITE_DCHECK_EQ(concat_size, output_shape.Dims(axis));
+  int64_t outer_size = 1;
+  for (int i = 0; i < axis; ++i) {
+    outer_size *= output_shape.Dims(i);
+  }
+  // For all input arrays,
+  // FlatSize() = outer_size * Dims(axis) * base_inner_size;
+  int64_t base_inner_size = 1;
+  for (int i = axis + 1; i < concat_dimensions; ++i) {
+    base_inner_size *= output_shape.Dims(i);
+  }
+
+  Scalar* output_ptr = output_data;
+  for (int k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int copy_size = input_shapes[i]->Dims(axis) * base_inner_size;
+      const Scalar* input_ptr = input_data[i] + k * copy_size;
+      memcpy(output_ptr, input_ptr, copy_size * sizeof(Scalar));
+      output_ptr += copy_size;
+    }
+  }
+}
+
+// TODO(b/174275780): The quantized implementation of concatentation isn't fully
+// quantized as it takes scale as a floating point value. This should be fixed
+// when optimizng this routine further.
+inline void ConcatenationWithScaling(const ConcatenationParams& params,
+                                     const RuntimeShape* const* input_shapes,
+                                     const uint8_t* const* input_data,
+                                     const RuntimeShape& output_shape,
+                                     uint8_t* output_data) {
+  int axis = params.axis;
+  const int32_t* input_zeropoint = params.input_zeropoint;
+  const float* input_scale = params.input_scale;
+  int inputs_count = params.inputs_count;
+  const int32_t output_zeropoint = params.output_zeropoint;
+  const float output_scale = params.output_scale;
+
+  const int concat_dimensions = output_shape.DimensionsCount();
+  TFLITE_DCHECK_LT(axis, concat_dimensions);
+
+  int64_t concat_size = 0;
+  for (int i = 0; i < inputs_count; i++) {
+    TFLITE_DCHECK_EQ(input_shapes[i]->DimensionsCount(), concat_dimensions);
+    for (int j = 0; j < concat_dimensions; j++) {
+      if (j != axis) {
+        MatchingDim(*input_shapes[i], j, output_shape, j);
+      }
+    }
+    concat_size += input_shapes[i]->Dims(axis);
+  }
+  TFLITE_DCHECK_EQ(concat_size, output_shape.Dims(axis));
+  int64_t outer_size = 1;
+  for (int i = 0; i < axis; ++i) {
+    outer_size *= output_shape.Dims(i);
+  }
+  // For all input arrays,
+  // FlatSize() = outer_size * Dims(axis) * base_inner_size;
+  int64_t base_inner_size = 1;
+  for (int i = axis + 1; i < concat_dimensions; ++i) {
+    base_inner_size *= output_shape.Dims(i);
+  }
+
+  const float inverse_output_scale = 1.f / output_scale;
+  uint8_t* output_ptr = output_data;
+  for (int k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int copy_size = input_shapes[i]->Dims(axis) * base_inner_size;
+      const uint8_t* input_ptr = input_data[i] + k * copy_size;
+      if (input_zeropoint[i] == output_zeropoint &&
+          input_scale[i] == output_scale) {
+        memcpy(output_ptr, input_ptr, copy_size);
+      } else {
+        const float scale = input_scale[i] * inverse_output_scale;
+        const float bias = -input_zeropoint[i] * scale;
+        for (int j = 0; j < copy_size; ++j) {
+          const int32_t value = static_cast<int32_t>(tflite::TfLiteRound(
+                                    input_ptr[j] * scale + bias)) +
+                                output_zeropoint;
+          output_ptr[j] = static_cast<uint8_t>(
+              std::max<int32_t>(std::min<int32_t>(255, value), 0));
+        }
+      }
+      output_ptr += copy_size;
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/conv.h

@@ -0,0 +1,264 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void Conv(const ConvParams& params, const RuntimeShape& input_shape,
+                 const float* input_data, const RuntimeShape& filter_shape,
+                 const float* filter_data, const RuntimeShape& bias_shape,
+                 const float* bias_data, const RuntimeShape& output_shape,
+                 float* output_data, const RuntimeShape& im2col_shape,
+                 float* im2col_data) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          float total = 0.f;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                float input_value = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                float filter_value = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                total += (input_value * filter_value);
+              }
+            }
+          }
+          float bias_value = 0.0f;
+          if (bias_data) {
+            bias_value = bias_data[out_channel];
+          }
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              ActivationFunctionWithMinMax(total + bias_value,
+                                           output_activation_min,
+                                           output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void Conv(const ConvParams& params, const RuntimeShape& input_shape,
+                 const uint8_t* input_data, const RuntimeShape& filter_shape,
+                 const uint8_t* filter_data, const RuntimeShape& bias_shape,
+                 const int32_t* bias_data, const RuntimeShape& output_shape,
+                 uint8_t* output_data, const RuntimeShape& im2col_shape,
+                 uint8_t* im2col_data, void* cpu_backend_context) {
+  (void)cpu_backend_context;  // only used in optimized code.
+  (void)im2col_data;          // only used in optimized code.
+  (void)im2col_shape;         // only used in optimized code.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          int32_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                int32_t input_val = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                acc +=
+                    (filter_val + filter_offset) * (input_val + input_offset);
+              }
+            }
+          }
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          acc = MultiplyByQuantizedMultiplier(acc, output_multiplier,
+                                              output_shift);
+          acc += output_offset;
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<uint8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+inline void HybridConvPerChannel(
+    const ConvParams& params, float* scaling_factors_ptr,
+    const RuntimeShape& input_shape, const int8_t* input_data,
+    const RuntimeShape& filter_shape, const int8_t* filter_data,
+    const RuntimeShape& bias_shape, const float* bias_data,
+    const RuntimeShape& output_shape, float* output_data,
+    const RuntimeShape& im2col_shape, int8_t* im2col_data,
+    const float* per_channel_scale, int32_t* input_offset) {
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          const int in_x_origin = (out_x * stride_width) - pad_width;
+          const int in_y_origin = (out_y * stride_height) - pad_height;
+          int32_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // If the location is outside the bounds of the input image,
+                // use zero as a default value.
+                if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height)) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  acc += filter_val * (input_val - input_offset[batch]);
+                }
+              }
+            }
+          }
+          float acc_float =
+              acc * per_channel_scale[out_channel] * scaling_factors_ptr[batch];
+          if (bias_data) {
+            acc_float += bias_data[out_channel];
+          }
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              ActivationFunctionWithMinMax(acc_float, output_activation_min,
+                                           output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/cumsum.h

@@ -0,0 +1,175 @@
+/* Copyright 2021 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_KERNELS_INTERNAL_REFERENCE_CUMSUM_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CUMSUM_H_
+
+#include <algorithm>
+#include <cstdint>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+inline void CumSum(const T* input_data, const RuntimeShape& shape, int32_t axis,
+                   bool exclusive, bool reverse, T* output_data) {
+  const int32_t rank = shape.DimensionsCount();
+  TFLITE_DCHECK_GE(rank, 1);
+  TFLITE_DCHECK_GE(axis, 0);
+  TFLITE_DCHECK_LT(axis, rank);
+
+  size_t inner = 1;
+  size_t outer = 1;
+  size_t depth = 1;
+  for (int32_t i = 0; i < rank; i++) {
+    if (i < axis)
+      inner *= shape.Dims(i);
+    else if (i > axis)
+      outer *= shape.Dims(i);
+    else
+      depth = shape.Dims(i);
+  }
+
+  for (size_t outer_index = 0; outer_index < outer; outer_index++) {
+    size_t outer_index_adj;
+    if (reverse)
+      outer_index_adj = (outer - 1) - outer_index;
+    else
+      outer_index_adj = outer_index;
+    for (size_t inner_index = 0; inner_index < inner; inner_index++) {
+      T accumulator = 0;
+      size_t inner_index_adj;
+      if (reverse)
+        inner_index_adj = (inner - 1) - inner_index;
+      else
+        inner_index_adj = inner_index;
+      for (size_t depth_index = 0; depth_index < depth; depth_index++) {
+        size_t depth_index_adj;
+        if (reverse)
+          depth_index_adj = (depth - 1) - depth_index;
+        else
+          depth_index_adj = depth_index;
+
+        size_t index = outer_index_adj;
+        index += inner_index_adj * depth * outer;
+        index += depth_index_adj * outer;
+
+        if (exclusive) {
+          output_data[index] = accumulator;
+          accumulator += input_data[index];
+        } else {
+          accumulator += input_data[index];
+          output_data[index] = accumulator;
+        }
+      }
+    }
+  }
+}
+
+//
+// Quantized INT8 CUMSUM
+//
+inline void CumSum(const ArithmeticParams& params, const int8_t* input_data,
+                   const RuntimeShape& shape, int32_t axis, bool exclusive,
+                   bool reverse, int8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  // Input offset is negative input zero point. Activation tensors are
+  // asymmetric quantized so they span the full int8 range.
+  // All inputs should have same zero-point and scale, this is checked during
+  // Prepare stage.
+  TFLITE_DCHECK_GE(-params.input1_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_LE(-params.input1_offset, std::numeric_limits<int8_t>::max());
+
+  const int32_t rank = shape.DimensionsCount();
+  TFLITE_DCHECK_GE(rank, 1);
+  TFLITE_DCHECK_GE(axis, 0);
+  TFLITE_DCHECK_LT(axis, rank);
+
+  size_t inner = 1;
+  size_t outer = 1;
+  size_t depth = 1;
+  for (int32_t i = 0; i < rank; i++) {
+    if (i < axis)
+      inner *= shape.Dims(i);
+    else if (i > axis)
+      outer *= shape.Dims(i);
+    else
+      depth = shape.Dims(i);
+  }
+
+  for (size_t outer_index = 0; outer_index < outer; outer_index++) {
+    size_t outer_index_adj;
+    if (reverse)
+      outer_index_adj = (outer - 1) - outer_index;
+    else
+      outer_index_adj = outer_index;
+    for (size_t inner_index = 0; inner_index < inner; inner_index++) {
+      int32_t accumulator = params.input1_offset;  // accumulator = 0
+      accumulator *= (1 << params.left_shift);
+      accumulator = MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          accumulator, params.input1_multiplier, params.input1_shift);
+
+      size_t inner_index_adj;
+      if (reverse)
+        inner_index_adj = (inner - 1) - inner_index;
+      else
+        inner_index_adj = inner_index;
+
+      for (size_t depth_index = 0; depth_index < depth; depth_index++) {
+        size_t depth_index_adj;
+        if (reverse)
+          depth_index_adj = (depth - 1) - depth_index;
+        else
+          depth_index_adj = depth_index;
+
+        size_t index = outer_index_adj;
+        index += inner_index_adj * depth * outer;
+        index += depth_index_adj * outer;
+
+        const int32_t y = params.input1_offset + input_data[index];
+        const int32_t shifted_y = y * (1 << params.left_shift);
+        const int32_t scaled_y = MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_y, params.input1_multiplier, params.input1_shift);
+
+        int32_t scaled_output;
+        if (exclusive) {
+          scaled_output = accumulator;
+          accumulator += scaled_y;
+        } else {
+          accumulator += scaled_y;
+          scaled_output = accumulator;
+        }
+
+        const int32_t raw_output =
+            MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                scaled_output, params.output_multiplier, params.output_shift) +
+            params.output_offset;
+        const int32_t clamped_output =
+            std::min(params.quantized_activation_max,
+                     std::max(params.quantized_activation_min, raw_output));
+        output_data[index] = static_cast<int8_t>(clamped_output);
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CUMSUM_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/depth_to_space.h

@@ -0,0 +1,79 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_DEPTH_TO_SPACE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTH_TO_SPACE_H_
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+inline void DepthToSpace(const tflite::DepthToSpaceParams& op_params,
+                         const RuntimeShape& unextended_input_shape,
+                         const T* input_data,
+                         const RuntimeShape& unextended_output_shape,
+                         T* output_data) {
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  const int input_depth = input_shape.Dims(3);
+  const int input_width = input_shape.Dims(2);
+  const int input_height = input_shape.Dims(1);
+  const int input_batch = input_shape.Dims(0);
+
+  const int output_depth = output_shape.Dims(3);
+  const int output_width = output_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_batch = output_shape.Dims(0);
+
+  const int32_t block_size = op_params.block_size;
+
+  TFLITE_DCHECK_EQ(input_width * block_size, output_width);
+  TFLITE_DCHECK_EQ(input_height * block_size, output_height);
+  TFLITE_DCHECK_EQ(input_depth, output_depth * block_size * block_size);
+  TFLITE_DCHECK_EQ(input_batch, output_batch);
+
+  for (int out_b = 0; out_b < output_batch; ++out_b) {
+    for (int out_h = 0; out_h < output_height; ++out_h) {
+      for (int out_w = 0; out_w < output_width; ++out_w) {
+        for (int out_d = 0; out_d < output_depth; ++out_d) {
+          const int in_d =
+              out_d + ((out_h % block_size) * block_size + out_w % block_size) *
+                          output_depth;
+
+          const int in_w = out_w / block_size;
+          const int in_h = out_h / block_size;
+          const int in_b = out_b;
+
+          const int input_index = Offset(input_shape, in_b, in_h, in_w, in_d);
+          const int output_index =
+              Offset(output_shape, out_b, out_h, out_w, out_d);
+
+          output_data[output_index] = input_data[input_index];
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTH_TO_SPACE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h

@@ -0,0 +1,100 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void DepthwiseConv(
+    const DepthwiseParams& params, const RuntimeShape& input_shape,
+    const float* input_data, const RuntimeShape& filter_shape,
+    const float* filter_data, const RuntimeShape& bias_shape,
+    const float* bias_data, const RuntimeShape& output_shape,
+    float* output_data) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+  for (int b = 0; b < batches; ++b) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int ic = 0; ic < input_depth; ++ic) {
+          for (int m = 0; m < depth_multiplier; m++) {
+            const int oc = m + ic * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            float total = 0.f;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // If the location is outside the bounds of the input image,
+                // use zero as a default value.
+                if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height)) {
+                  float input_value =
+                      input_data[Offset(input_shape, b, in_y, in_x, ic)];
+                  float filter_value = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, oc)];
+                  total += (input_value * filter_value);
+                }
+              }
+            }
+            float bias_value = 0.0f;
+            if (bias_data) {
+              bias_value = bias_data[oc];
+            }
+            output_data[Offset(output_shape, b, out_y, out_x, oc)] =
+                ActivationFunctionWithMinMax(total + bias_value,
+                                             output_activation_min,
+                                             output_activation_max);
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // end namespace reference_ops
+}  // end namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h

@@ -0,0 +1,319 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_
+
+#include <algorithm>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+// Used in tests and template parameters to control which version of depthwise
+// convolution is called. Primarily for reference code, and specializations
+// forced in tests.
+enum class DepthwiseConvImplementation {
+  // Run all tests against kUseStandardEntry even if also testing another
+  // kernel, since we need to be sure that the main DepthwiseConv() function in
+  // optimized_ops.h dispatches to a correctly-executing kernel.
+  kNone = 0,                 // The "default" option: use the normal
+                             // DepthwiseConv kernel (entry) function.
+  kUseGenericKernel,         // Forced use of generic kernel.
+  kUseNeon3x3,               // 3x3 kernel that uses NEON when available.
+  kUseNeon3x3DotProduct,     // 3x3 kernel that uses dot-product enabled NEON
+                             // when available.
+  kUseCModel3x3DotProduct,   // 3x3 kernel, reference C model that is intended
+                             // to match overall design NEON code.
+  kUseUnwound3x3DotProduct,  // 3x3 kernel, reference C model with unwound loops
+                             // and some arrays.
+  kUseIntrinsics3x3DotProduct,  // 3x3 kernel using NEON intrinsics.
+};
+
+// Category of depthwise convolution output rounding.
+enum class DepthwiseConvOutputRounding {
+  kNone = 0,      // Invalid: specific method must be specified.
+  kAwayFromZero,  // Original method: exact halves rounded away from zero.
+  kUpward,        // Halves towards +infinity: adds 0.5 before truncate.
+  // This is where a future kNearestEven would be placed.
+};
+
+// Category of depthwise convolution depth multiplication.
+enum class DepthwiseConvDepthMultiplication {
+  kNoMultiplication = 0,  // Depth multiplier = 1.
+  kUnitInputDepth,        // Input depth = 1, output depth = depth multiplier.
+};
+
+namespace reference_ops {
+namespace depthwise_conv {
+
+template <DepthwiseConvOutputRounding output_rounding>
+inline int32_t DepthwiseConvRound(int32_t x, int32_t quantized_multiplier,
+                                  int shift) {
+  TFLITE_DCHECK_NE(output_rounding, DepthwiseConvOutputRounding::kNone);
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+
+// Single-rounding MultiplyByQuantizedMultiplier
+#if TFLITE_SINGLE_ROUNDING
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kAwayFromZero>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  using gemmlowp::RoundingDivideByPOT;
+  using gemmlowp::SaturatingRoundingDoublingHighMul;
+  int left_shift = shift > 0 ? shift : 0;
+  int right_shift = shift > 0 ? 0 : -shift;
+  return RoundingDivideByPOT(SaturatingRoundingDoublingHighMul(
+                                 x * (1 << left_shift), quantized_multiplier),
+                             right_shift);
+}
+
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kUpward>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+// Double-rounding MultiplyByQuantizedMultiplier
+#else
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kAwayFromZero>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kUpward>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  using gemmlowp::SaturatingRoundingDoublingHighMul;
+  const int left_shift = shift > 0 ? shift : 0;
+  const int right_shift = shift > 0 ? 0 : -shift;
+  const int rounding_offset = right_shift > 0 ? 1 << (right_shift - 1) : 0;
+  return (SaturatingRoundingDoublingHighMul(x * (1 << left_shift),
+                                            quantized_multiplier) +
+          rounding_offset) >>
+         right_shift;
+}
+#endif  // TFLITE_SINGLE_ROUNDING
+
+template <DepthwiseConvOutputRounding output_rounding>
+struct DepthwiseConvBasicKernel {
+  static inline void Run(
+      const DepthwiseParams& params, const RuntimeShape& input_shape,
+      const uint8_t* input_data, const RuntimeShape& filter_shape,
+      const uint8_t* filter_data, const RuntimeShape& bias_shape,
+      const int32_t* bias_data, const RuntimeShape& output_shape,
+      uint8_t* output_data) {
+    const int stride_width = params.stride_width;
+    const int stride_height = params.stride_height;
+    const int dilation_width_factor = params.dilation_width_factor;
+    const int dilation_height_factor = params.dilation_height_factor;
+    const int pad_width = params.padding_values.width;
+    const int pad_height = params.padding_values.height;
+    const int depth_multiplier = params.depth_multiplier;
+    const int32_t output_activation_min = params.quantized_activation_min;
+    const int32_t output_activation_max = params.quantized_activation_max;
+    const int32_t input_offset = params.input_offset;
+    const int32_t filter_offset = params.weights_offset;
+    const int32_t output_offset = params.output_offset;
+    const int32_t output_multiplier = params.output_multiplier;
+    const int output_shift = params.output_shift;
+    TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+    TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+    const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+    const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+    const int input_height = input_shape.Dims(1);
+    const int input_width = input_shape.Dims(2);
+    const int input_depth = input_shape.Dims(3);
+    const int filter_height = filter_shape.Dims(1);
+    const int filter_width = filter_shape.Dims(2);
+    const int output_height = output_shape.Dims(1);
+    const int output_width = output_shape.Dims(2);
+    TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+    for (int b = 0; b < batches; ++b) {
+      for (int out_y = 0; out_y < output_height; ++out_y) {
+        for (int out_x = 0; out_x < output_width; ++out_x) {
+          for (int ic = 0; ic < input_depth; ++ic) {
+            for (int m = 0; m < depth_multiplier; m++) {
+              const int oc = m + ic * depth_multiplier;
+              const int in_x_origin = (out_x * stride_width) - pad_width;
+              const int in_y_origin = (out_y * stride_height) - pad_height;
+              int32_t acc = 0;
+              for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+                for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                  const int in_x =
+                      in_x_origin + dilation_width_factor * filter_x;
+                  const int in_y =
+                      in_y_origin + dilation_height_factor * filter_y;
+                  // If the location is outside the bounds of the input image,
+                  // use zero as a default value.
+                  if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                      (in_y < input_height)) {
+                    int32_t input_val =
+                        input_data[Offset(input_shape, b, in_y, in_x, ic)];
+                    int32_t filter_val = filter_data[Offset(
+                        filter_shape, 0, filter_y, filter_x, oc)];
+                    acc += (filter_val + filter_offset) *
+                           (input_val + input_offset);
+                  }
+                }
+              }
+              if (bias_data) {
+                acc += bias_data[oc];
+              }
+              acc = DepthwiseConvRound<output_rounding>(acc, output_multiplier,
+                                                        output_shift);
+              acc += output_offset;
+              acc = std::max(acc, output_activation_min);
+              acc = std::min(acc, output_activation_max);
+              output_data[Offset(output_shape, b, out_y, out_x, oc)] =
+                  static_cast<uint8_t>(acc);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  // TODO(b/148596273): Reconcile reference versions, perhaps with common
+  // MultiplyByQuantizedMultiplier or DepthwiseConvRound function.
+  static inline void RunPerChannel(
+      const DepthwiseParams& params, const RuntimeShape& input_shape,
+      const int8_t* input_data, const RuntimeShape& filter_shape,
+      const int8_t* filter_data, const RuntimeShape& bias_shape,
+      const int32_t* bias_data, const RuntimeShape& output_shape,
+      int8_t* output_data) {
+    // Get parameters.
+    // TODO(b/141565753): Re-introduce ScopedProfilingLabel on Micro.
+    const int stride_width = params.stride_width;
+    const int stride_height = params.stride_height;
+    const int dilation_width_factor = params.dilation_width_factor;
+    const int dilation_height_factor = params.dilation_height_factor;
+    const int pad_width = params.padding_values.width;
+    const int pad_height = params.padding_values.height;
+    const int depth_multiplier = params.depth_multiplier;
+    const int32_t input_offset = params.input_offset;
+    const int32_t output_offset = params.output_offset;
+    const int32_t output_activation_min = params.quantized_activation_min;
+    const int32_t output_activation_max = params.quantized_activation_max;
+    const int32_t* output_multiplier = params.output_multiplier_per_channel;
+    const int32_t* output_shift = params.output_shift_per_channel;
+
+    // Check dimensions of the tensors.
+    TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+    TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+    const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+    const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+    const int input_height = input_shape.Dims(1);
+    const int input_width = input_shape.Dims(2);
+    const int input_depth = input_shape.Dims(3);
+    const int filter_height = filter_shape.Dims(1);
+    const int filter_width = filter_shape.Dims(2);
+    const int output_height = output_shape.Dims(1);
+    const int output_width = output_shape.Dims(2);
+    TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+    for (int batch = 0; batch < batches; ++batch) {
+      for (int out_y = 0; out_y < output_height; ++out_y) {
+        for (int out_x = 0; out_x < output_width; ++out_x) {
+          for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+            for (int m = 0; m < depth_multiplier; ++m) {
+              const int output_channel = m + in_channel * depth_multiplier;
+              const int in_x_origin = (out_x * stride_width) - pad_width;
+              const int in_y_origin = (out_y * stride_height) - pad_height;
+              int32_t acc = 0;
+              for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+                for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                  const int in_x =
+                      in_x_origin + dilation_width_factor * filter_x;
+                  const int in_y =
+                      in_y_origin + dilation_height_factor * filter_y;
+                  // Zero padding by omitting the areas outside the image.
+                  const bool is_point_inside_image =
+                      (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                      (in_y < input_height);
+                  if (is_point_inside_image) {
+                    int32_t input_val = input_data[Offset(
+                        input_shape, batch, in_y, in_x, in_channel)];
+                    int32_t filter_val = filter_data[Offset(
+                        filter_shape, 0, filter_y, filter_x, output_channel)];
+                    // Accumulate with 32 bits accumulator.
+                    // In the nudging process during model quantization, we
+                    // force real value of 0.0 be represented by a quantized
+                    // value. This guarantees that the input_offset is a int8_t,
+                    // even though it is represented using int32_t. int32_t +=
+                    // int8_t
+                    // * (int8_t - int8_t) so the highest value we can get from
+                    // each accumulation is [-127, 127] * ([-128, 127] -
+                    // [-128, 127]), which is [-32512, 32512]. log2(32512)
+                    // = 14.98, which means we can accumulate at least 2^16
+                    // multiplications without overflow. The accumulator is
+                    // applied to a filter so the accumulation logic will hold
+                    // as long as the filter size (filter_y * filter_x *
+                    // in_channel) does not exceed 2^16, which is the case in
+                    // all the models we have seen so far.
+                    acc += filter_val * (input_val + input_offset);
+                  }
+                }
+              }
+              if (bias_data) {
+                acc += bias_data[output_channel];
+              }
+              acc = DepthwiseConvRound<output_rounding>(
+                  acc, output_multiplier[output_channel],
+                  output_shift[output_channel]);
+              acc += output_offset;
+              acc = std::max(acc, output_activation_min);
+              acc = std::min(acc, output_activation_max);
+              output_data[Offset(output_shape, batch, out_y, out_x,
+                                 output_channel)] = static_cast<int8_t>(acc);
+            }
+          }
+        }
+      }
+    }
+  }
+};
+
+}  // namespace depthwise_conv
+
+inline void DepthwiseConv(
+    const DepthwiseParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& filter_shape,
+    const uint8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    uint8_t* output_data) {
+  return depthwise_conv::DepthwiseConvBasicKernel<
+      DepthwiseConvOutputRounding::kAwayFromZero>::Run(params, input_shape,
+                                                       input_data, filter_shape,
+                                                       filter_data, bias_shape,
+                                                       bias_data, output_shape,
+                                                       output_data);
+}
+
+}  // namespace reference_ops
+}  // end namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/dequantize.h

@@ -0,0 +1,78 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_
+
+#include <limits.h>
+
+#include <vector>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Dequantizes into a float without rounding.
+template <typename InputT, typename OutputT>
+inline void Dequantize(const tflite::DequantizationParams& op_params,
+                       const RuntimeShape& input_shape,
+                       const InputT* input_data,
+                       const RuntimeShape& output_shape, OutputT* output_data) {
+  int32_t zero_point = op_params.zero_point;
+  const double scale = op_params.scale;
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    const int32_t val = input_data[i];
+    const OutputT result = static_cast<OutputT>(scale * (val - zero_point));
+    output_data[i] = result;
+  }
+}
+
+// Dequantizes per-channel quantized tensor to float.
+template <typename T>
+inline void PerChannelDequantize(
+    const tflite::PerChannelDequantizationParams& op_params,
+    const RuntimeShape& input_shape, const T* input_data,
+    const RuntimeShape& output_shape, float* output_data) {
+  // Ensure flat size is same.
+  MatchingFlatSize(input_shape, output_shape);
+
+  const int32_t* zero_point = op_params.zero_point;
+  const float* scale = op_params.scale;
+  const int32_t quantized_dimension = op_params.quantized_dimension;
+  const int32_t num_dims = input_shape.DimensionsCount();
+  const int32_t* dims_data = input_shape.DimsData();
+  std::vector<int> current_dim(num_dims, 0);
+
+  do {
+    size_t offset =
+        ReducedOutputOffset(num_dims, reinterpret_cast<const int*>(dims_data),
+                            current_dim.data(), 0, nullptr);
+    const int channel = current_dim[quantized_dimension];
+    const int32_t val = input_data[offset];
+    const float result =
+        static_cast<float>(scale[channel] * (val - zero_point[channel]));
+    output_data[offset] = result;
+  } while (NextIndex(num_dims, reinterpret_cast<const int*>(dims_data),
+                     current_dim.data()));
+}
+
+}  // namespace reference_ops
+
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/elu.h

@@ -0,0 +1,37 @@
+/* Copyright 2021 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_KERNELS_INTERNAL_REFERENCE_ELU_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ELU_H_
+
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void Elu(const RuntimeShape& input_shape, const float* input_data,
+                const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    const float val = input_data[i];
+    output_data[i] = val < 0.0f ? TfLiteExpm1(val) : val;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ELU_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/exp.h

@@ -0,0 +1,38 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_EXP_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_EXP_H_
+
+#include <cmath>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+inline void Exp(const T* input_data, const size_t num_elements,
+                T* output_data) {
+  ruy::profiler::ScopeLabel label("Exp");
+  for (size_t idx = 0; idx < num_elements; ++idx) {
+    output_data[idx] = std::exp(input_data[idx]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_EXP_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/fill.h

@@ -0,0 +1,38 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_FILL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FILL_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+void Fill(const RuntimeShape& value_shape, const T* value_data,
+          const RuntimeShape& output_shape, T* output_data) {
+  TFLITE_DCHECK_EQ(value_shape.DimensionsCount(), 0);
+  const int flat_size = output_shape.FlatSize();
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = *value_data;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FILL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/floor.h

@@ -0,0 +1,39 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_FLOOR_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void Floor(const RuntimeShape& input_shape, const float* input_data,
+                  const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    int offset = i;
+    output_data[offset] = std::floor(input_data[offset]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/floor_div.h

@@ -0,0 +1,35 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_FLOOR_DIV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_DIV_H_
+
+#include <cmath>
+#include <functional>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+T FloorDiv(T input1, T input2) {
+  return std::floor(std::divides<double>()(static_cast<double>(input1),
+                                           static_cast<double>(input2)));
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_DIV_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/floor_mod.h

@@ -0,0 +1,44 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_FLOOR_MOD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_MOD_H_
+
+#include <cmath>
+#include <functional>
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+T FloorMod(T input1, T input2) {
+  struct FloatMod {
+    float operator()(const float lhs, const float rhs) const {
+      return std::fmod(lhs, rhs);
+    }
+  };
+  using ModFunc = typename std::conditional<std::is_integral<T>::value,
+                                            std::modulus<T>, FloatMod>::type;
+  ModFunc mod_func;
+  T trunc_mod = mod_func(input1, input2);
+  return (trunc_mod != 0) && ((input2 < 0) != (trunc_mod < 0))
+             ? (trunc_mod + input2)
+             : trunc_mod;
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_MOD_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/fully_connected.h

@@ -0,0 +1,321 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const float* input_data, const RuntimeShape& weights_shape,
+    const float* weights_data, const RuntimeShape& bias_shape,
+    const float* bias_data, const RuntimeShape& output_shape,
+    float* output_data) {
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dims_count = output_shape.DimensionsCount();
+  const int weights_dims_count = weights_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dims_count - 1);
+  const int output_depth = MatchingDim(weights_shape, weights_dims_count - 2,
+                                       output_shape, output_dims_count - 1);
+  const int accum_depth = weights_shape.Dims(weights_dims_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      float total = 0.f;
+      for (int d = 0; d < accum_depth; ++d) {
+        total += input_data[b * accum_depth + d] *
+                 weights_data[out_c * accum_depth + d];
+      }
+      float bias_value = 0.0f;
+      if (bias_data) {
+        bias_value = bias_data[out_c];
+      }
+      output_data[out_c + output_depth * b] = ActivationFunctionWithMinMax(
+          total + bias_value, output_activation_min, output_activation_max);
+    }
+  }
+}
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& filter_shape,
+    const uint8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    uint8_t* output_data) {
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = MatchingDim(filter_shape, filter_dim_count - 2,
+                                       output_shape, output_dim_count - 1);
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      int32_t acc = 0;
+      for (int d = 0; d < accum_depth; ++d) {
+        int32_t input_val = input_data[b * accum_depth + d];
+        int32_t filter_val = filter_data[out_c * accum_depth + d];
+        acc += (filter_val + filter_offset) * (input_val + input_offset);
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc += output_offset;
+      acc = std::max(acc, output_activation_min);
+      acc = std::min(acc, output_activation_max);
+      output_data[out_c + output_depth * b] = static_cast<uint8_t>(acc);
+    }
+  }
+}
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& filter_shape,
+    const uint8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(output_offset, 0);
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = MatchingDim(filter_shape, filter_dim_count - 2,
+                                       output_shape, output_dim_count - 1);
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      // Internal accumulation.
+      // Initialize accumulator with the bias-value.
+      int32_t accum = bias_data[out_c];
+      // Accumulation loop.
+      for (int d = 0; d < accum_depth; ++d) {
+        int16_t input_val = input_data[b * accum_depth + d] + input_offset;
+        int16_t filter_val =
+            filter_data[out_c * accum_depth + d] + filter_offset;
+        accum += filter_val * input_val;
+      }
+      // Down-scale the final int32_t accumulator to the scale used by our
+      // (16-bit, typically 3 integer bits) fixed-point format. The quantized
+      // multiplier and shift here have been pre-computed offline
+      // (e.g. by toco).
+      accum =
+          MultiplyByQuantizedMultiplier(accum, output_multiplier, output_shift);
+      // Saturate, cast to int16_t, and store to output array.
+      accum = std::max(accum, output_activation_min - output_offset);
+      accum = std::min(accum, output_activation_max - output_offset);
+      accum += output_offset;
+      output_data[out_c + output_depth * b] = accum;
+    }
+  }
+}
+
+inline void ShuffledFullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& weights_shape,
+    const uint8_t* shuffled_weights_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data, uint8_t* shuffled_input_workspace_data) {
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  TFLITE_DCHECK_GE(input_shape.DimensionsCount(), 1);
+  TFLITE_DCHECK_GE(weights_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1);
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int weights_dim_count = weights_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = MatchingDim(weights_shape, weights_dim_count - 2,
+                                       output_shape, output_dim_count - 1);
+  const int accum_depth = weights_shape.Dims(weights_dim_count - 1);
+  TFLITE_DCHECK((accum_depth % 16) == 0);
+  TFLITE_DCHECK((output_depth % 4) == 0);
+
+  // Shuffling and xoring of input activations into the workspace buffer
+  uint8_t* shuffled_input_workspace_ptr = shuffled_input_workspace_data;
+  if (batches == 1) {
+    for (int i = 0; i < accum_depth; i++) {
+      shuffled_input_workspace_data[i] = input_data[i] ^ 0x80;
+    }
+  } else if (batches == 4) {
+    for (int c = 0; c < accum_depth; c += 16) {
+      for (int b = 0; b < 4; b++) {
+        const uint8_t* src_data_ptr = input_data + b * accum_depth + c;
+        for (int j = 0; j < 16; j++) {
+          uint8_t src_val = *src_data_ptr++;
+          // Flip the sign bit, so that the kernel will only need to
+          // reinterpret these uint8_t values as int8_t, getting for free the
+          // subtraction of the zero_point value 128.
+          uint8_t dst_val = src_val ^ 0x80;
+          *shuffled_input_workspace_ptr++ = dst_val;
+        }
+      }
+    }
+  } else {
+    TFLITE_DCHECK(false);
+    return;
+  }
+
+  // Actual computation
+  if (batches == 1) {
+    int16_t* output_ptr = output_data;
+    // Shuffled weights have had their sign bit (0x80) pre-flipped (xor'd)
+    // so that just reinterpreting them as int8_t values is equivalent to
+    // subtracting 128 from them, thus implementing for free the subtraction of
+    // the zero_point value 128.
+    const int8_t* shuffled_weights_ptr =
+        reinterpret_cast<const int8_t*>(shuffled_weights_data);
+    // Likewise, we preshuffled and pre-xored the input data above.
+    const int8_t* shuffled_input_data =
+        reinterpret_cast<const int8_t*>(shuffled_input_workspace_data);
+    for (int c = 0; c < output_depth; c += 4) {
+      // Internal accumulation.
+      // Initialize accumulator with the bias-value.
+      int32_t accum[4] = {0};
+      // Accumulation loop.
+      for (int d = 0; d < accum_depth; d += 16) {
+        for (int i = 0; i < 4; i++) {
+          for (int j = 0; j < 16; j++) {
+            int8_t input_val = shuffled_input_data[d + j];
+            int8_t weights_val = *shuffled_weights_ptr++;
+            accum[i] += weights_val * input_val;
+          }
+        }
+      }
+      for (int i = 0; i < 4; i++) {
+        // Add bias value
+        int32_t acc = accum[i] + bias_data[c + i];
+        // Down-scale the final int32_t accumulator to the scale used by our
+        // (16-bit, typically 3 integer bits) fixed-point format. The quantized
+        // multiplier and shift here have been pre-computed offline
+        // (e.g. by toco).
+        acc =
+            MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+        // Saturate, cast to int16_t, and store to output array.
+        acc = std::max(acc, output_activation_min);
+        acc = std::min(acc, output_activation_max);
+        output_ptr[c + i] = acc;
+      }
+    }
+  } else if (batches == 4) {
+    int16_t* output_ptr = output_data;
+    // Shuffled weights have had their sign bit (0x80) pre-flipped (xor'd)
+    // so that just reinterpreting them as int8_t values is equivalent to
+    // subtracting 128 from them, thus implementing for free the subtraction of
+    // the zero_point value 128.
+    const int8_t* shuffled_weights_ptr =
+        reinterpret_cast<const int8_t*>(shuffled_weights_data);
+    // Likewise, we preshuffled and pre-xored the input data above.
+    const int8_t* shuffled_input_data =
+        reinterpret_cast<const int8_t*>(shuffled_input_workspace_data);
+    for (int c = 0; c < output_depth; c += 4) {
+      const int8_t* shuffled_input_ptr = shuffled_input_data;
+      // Accumulation loop.
+      // Internal accumulation.
+      // Initialize accumulator with the bias-value.
+      int32_t accum[4][4];
+      for (int i = 0; i < 4; i++) {
+        for (int b = 0; b < 4; b++) {
+          accum[i][b] = 0;
+        }
+      }
+      for (int d = 0; d < accum_depth; d += 16) {
+        for (int i = 0; i < 4; i++) {
+          for (int b = 0; b < 4; b++) {
+            for (int j = 0; j < 16; j++) {
+              int8_t input_val = shuffled_input_ptr[16 * b + j];
+              int8_t weights_val = shuffled_weights_ptr[16 * i + j];
+              accum[i][b] += weights_val * input_val;
+            }
+          }
+        }
+        shuffled_input_ptr += 64;
+        shuffled_weights_ptr += 64;
+      }
+      for (int i = 0; i < 4; i++) {
+        for (int b = 0; b < 4; b++) {
+          // Add bias value
+          int32_t acc = accum[i][b] + bias_data[c + i];
+          // Down-scale the final int32_t accumulator to the scale used by our
+          // (16-bit, typically 3 integer bits) fixed-point format. The
+          // quantized multiplier and shift here have been pre-computed offline
+          // (e.g. by toco).
+          acc = MultiplyByQuantizedMultiplier(acc, output_multiplier,
+                                              output_shift);
+          // Saturate, cast to int16_t, and store to output array.
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_ptr[b * output_depth + c + i] = acc;
+        }
+      }
+    }
+  } else {
+    TFLITE_DCHECK(false);
+    return;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/hard_swish.h

@@ -0,0 +1,166 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline int16_t SaturatingLeftShift(int16_t value, int amount) {
+  int32_t result = static_cast<int32_t>(value) * (1 << amount);
+  result = std::min<int32_t>(result, std::numeric_limits<int16_t>::max());
+  result = std::max<int32_t>(result, std::numeric_limits<int16_t>::min());
+  return result;
+}
+
+// Similar to ARM instruction SQDMULH.
+// Similar to gemmlowp::SaturatingRoundingDoublingHighMul except
+// rounding to zero instead of to nearest (SQRDMULH).
+inline std::int16_t SaturatingDoublingHighMul(std::int16_t a, std::int16_t b) {
+  bool overflow = a == b && a == std::numeric_limits<std::int16_t>::min();
+  std::int32_t a_32(a);
+  std::int32_t b_32(b);
+  std::int32_t ab_32 = a_32 * b_32;
+  std::int16_t ab_x2_high16 = static_cast<std::int16_t>((ab_32) / (1 << 15));
+  return overflow ? std::numeric_limits<std::int16_t>::max() : ab_x2_high16;
+}
+
+template <typename T>
+inline void HardSwish(const RuntimeShape& input_shape, const T* input_data,
+                      const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("ReferenceHardSwish/Float");
+  auto matching_size = MatchingFlatSize(input_shape, output_shape);
+  const T* in_end = input_data + matching_size;
+  for (; input_data < in_end; input_data++, output_data++) {
+    const float in = *input_data;
+    *output_data =
+        in * std::min(static_cast<T>(6), std::max(static_cast<T>(0), in + 3)) /
+        6;
+  }
+}
+
+template <typename T>
+inline void HardSwish(const HardSwishParams& params,
+                      const RuntimeShape& input_shape, const T* input_data,
+                      const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("ReferenceHardSwish/Quantized");
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    const int16_t input_value = input_data[i] - params.input_zero_point;
+    // Left-shift as much as we can without overflow/saturation to put
+    // significant bits in the high bits of our 16-bit fixedpoint values, so
+    // that fixed-point approximate computations below are as accurate as
+    // possible.
+    const int16_t input_value_on_hires_input_scale = input_value * (1 << 7);
+    // Compute the input value on essentially the output scale, just not
+    // right-shifted yet. This is the value that we'll use in the (x >= +3)
+    // case, and that in the general case we'll multiply against the "relu-ish"
+    // fixed-point multiplier in [0, 1].
+    const int16_t input_value_on_preshift_output_scale =
+        gemmlowp::SaturatingRoundingDoublingHighMul(
+            input_value_on_hires_input_scale,
+            params.output_multiplier_fixedpoint_int16);
+    // Now compute the "relu-ish multiplier". In the (-3 <= x <= +3) case, that
+    // is just an affine rescaling of x from [-3, 3] to [0, 1]. In the general
+    // case, it is just that plus saturation at the boundaries of [-3, 3].
+    // First, we rescale from [-3, 3] to [-1, 1], saturating.
+    // That is done by rescaling the input value with a fixed-point multiplier
+    // (reluish_multiplier_fixedpoint) and bit-shift such that we represent
+    // that input value on the scale where the real value 3.0f is represented
+    // by the quantized value 32768.  (+32768 is actually not representable as
+    // int16_t, so this saturates at +32767, and that is seen empirically to be
+    // a negligible contribution to numerical error/bias).
+    //
+    // This code is careful to correctly implement any magnitude of multiplier,
+    // involving either a right shift or a left shift, with correct saturation
+    // behavior in the left-shift case. This forces this code to be more
+    // complicated, but is necessary for real applications: a partially
+    // trained quantized MobileNet v3-small model that motivated this code
+    // exhibits some large [min, max] range boundaries, of the order of
+    // magnitude of 10 or 100 depending on layers.
+    //
+    // The next few lines are basically just an ordinary
+    // MultiplyByQuantizedMultiplier, except that we are more careful here
+    // about the fine details of saturation when left-shifting, because here
+    // overflow in left-shift is a common case, not an anomaly as
+    // MultiplyByQuantizedMultiplier assumes.
+    int16_t reluish_value = input_value_on_hires_input_scale;
+    // Shift left, saturating, as much as we can while ensuring that this
+    // saturation will not contribute to the result. That is, left shift amount
+    // reduced by 1.
+    if (params.reluish_multiplier_exponent > 0) {
+      reluish_value = SaturatingLeftShift(
+          reluish_value, params.reluish_multiplier_exponent - 1);
+    }
+    // Apply the fixed-point multiplier, dividing the value by a divisor
+    // ranging in [1, 2].
+    reluish_value = gemmlowp::SaturatingRoundingDoublingHighMul(
+        reluish_value, params.reluish_multiplier_fixedpoint_int16);
+    // Apply the last bit of left-shift. Thus, in the left-shifting case, if
+    // any saturation affects the result, it is happening here --- any
+    // saturation having occurred above is overwritten here, not affecting the
+    // result.
+    if (params.reluish_multiplier_exponent > 0) {
+      reluish_value = SaturatingLeftShift(reluish_value, 1);
+    }
+    // Shift right, in the right-shifting case.
+    if (params.reluish_multiplier_exponent < 0) {
+      reluish_value = gemmlowp::RoundingDivideByPOT(
+          reluish_value, -params.reluish_multiplier_exponent);
+    }
+    // At this point we have rescaled the value into a 16bit fixedpoint
+    // reluish_value in [-1, 1].
+    // We now convert that to a 16bit fixedpoint value in [0, 1].
+    reluish_value = (reluish_value + (1 << 15)) >> 1;
+    // Use of SaturatingDoublingHighMul here is important to cancel the biases
+    // from the above SaturatingRoundingDoublingHighMul.
+    //
+    // On a partially trained MobileNet-v3-small,
+    //
+    //                                       | bias on    |  ImageNet
+    //                                       | quantized  |  Top-1
+    // Operation used here                   | values     |  accuracy (50k)
+    // --------------------------------------+------------+-----------
+    // SaturatingDoublingHighMul             | -0.0024    |  58.920
+    // SaturatingRoundingDoublingHighMul     | -0.0067    |  58.064
+    //
+    // In activations_test, this is covered by this testcase:
+    //     QuantizedActivationsOpTest.HardSwishBias
+    //
+    const int16_t preshift_output_value = SaturatingDoublingHighMul(
+        reluish_value, input_value_on_preshift_output_scale);
+    // We were so far operating on the pre-shift output scale. Now we finally
+    // apply that output shift, arriving at the final output scale.
+    int16_t output_value = gemmlowp::RoundingDivideByPOT(
+        preshift_output_value, -params.output_multiplier_exponent);
+    output_value += params.output_zero_point;
+    output_value =
+        std::min<int16_t>(output_value, std::numeric_limits<T>::max());
+    output_value =
+        std::max<int16_t>(output_value, std::numeric_limits<T>::min());
+    output_data[i] = output_value;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/add.h

@@ -0,0 +1,144 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_
+
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void CheckArithmeticParams(const ArithmeticParams& params) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  // Input offset is negative input zero point. Activation tensors are
+  // asymmetric quantized so they span the full int8 range.
+  TFLITE_DCHECK_GE(-params.input1_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_GE(-params.input2_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_LE(-params.input1_offset, std::numeric_limits<int8_t>::max());
+  TFLITE_DCHECK_LE(-params.input2_offset, std::numeric_limits<int8_t>::max());
+}
+
+inline void ElementWise(
+    int size, const ArithmeticParams& params, const int8_t* input1_data,
+    const int8_t* input2_data, int8_t* output_data,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    int8_t (*binary_func)(int8_t, int8_t, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[i], input2_data[i], params);
+  }
+}
+
+inline void BroadcastBinaryFunction4DSlow(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const int8_t* input1_data, const RuntimeShape& input2_shape,
+    const int8_t* input2_data, const RuntimeShape& output_shape,
+    int8_t* output_data,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    int8_t (*binary_func)(int8_t, int8_t, const ArithmeticParams&)) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          output_data[Offset(extended_output_shape, b, y, x, c)] = binary_func(
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)],
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)], params);
+        }
+      }
+    }
+  }
+}
+
+inline int8_t AddFunc(int8_t x, int8_t y, const ArithmeticParams& params) {
+  const int32_t input1_val = params.input1_offset + x;
+  const int32_t input2_val = params.input2_offset + y;
+  const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+  const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+  const int32_t scaled_input1_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input1_val, params.input1_multiplier, params.input1_shift);
+  const int32_t scaled_input2_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input2_val, params.input2_multiplier, params.input2_shift);
+  const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+  const int32_t raw_output =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          raw_sum, params.output_multiplier, params.output_shift) +
+      params.output_offset;
+  const int32_t clamped_output =
+      std::min(params.quantized_activation_max,
+               std::max(params.quantized_activation_min, raw_output));
+  return static_cast<int8_t>(clamped_output);
+}
+
+// Element-wise add that can often be used for inner loop of broadcast add as
+// well as the non-broadcast add.
+inline void AddElementwise(int size, const ArithmeticParams& params,
+                           const int8_t* input1_data, const int8_t* input2_data,
+                           int8_t* output_data) {
+  ElementWise(size, params, input1_data, input2_data, output_data,
+              CheckArithmeticParams, AddFunc);
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int8_t* input1_data,
+                const RuntimeShape& input2_shape, const int8_t* input2_data,
+                const RuntimeShape& output_shape, int8_t* output_data) {
+  CheckArithmeticParams(params);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  AddElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void BroadcastAdd4DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const int8_t* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const int8_t* input2_data,
+                               const RuntimeShape& output_shape,
+                               int8_t* output_data) {
+  BroadcastBinaryFunction4DSlow(params, input1_shape, input1_data, input2_shape,
+                                input2_data, output_shape, output_data,
+                                CheckArithmeticParams, AddFunc);
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h

@@ -0,0 +1,221 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+// Fixed-point per-channel-quantization convolution reference kernel.
+inline void ConvPerChannel(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data) {
+  // Get parameters.
+  const int32_t input_offset = params.input_offset;  // r = s(q - Z)
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int32_t output_offset = params.output_offset;
+
+  // Set min and max value of the output.
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Consistency check.
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  // Check dimensions of the tensors.
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          int32_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                int32_t input_val = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                // Accumulate with 32 bits accumulator.
+                // In the nudging process during model quantization, we force
+                // real value of 0.0 be represented by a quantized value. This
+                // guarantees that the input_offset is a int8_t, even though
+                // it is represented using int32_t. int32_t += int8_t *
+                // (int8_t - int8_t) so the highest value we can get from each
+                // accumulation is [-127, 127] * ([-128, 127] -
+                // [-128, 127]), which is [-32512, 32512]. log2(32512)
+                // = 14.98, which means we can accumulate at least 2^16
+                // multiplications without overflow. The accumulator is
+                // applied to a filter so the accumulation logic will hold as
+                // long as the filter size (filter_y * filter_x * in_channel)
+                // does not exceed 2^16, which is the case in all the models
+                // we have seen so far.
+                // TODO(b/174275578): Add a check to make sure the
+                // accumulator depth is smaller than 2^16.
+                acc += filter_val * (input_val + input_offset);
+              }
+            }
+          }
+
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          acc += output_offset;
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+// Fixed-point per-channel-quantization convolution reference kernel.
+// 16-bit data and 8-bit filter
+inline void ConvPerChannel(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const std::int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  // Get parameters.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+
+  // Set min and max value of the output.
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Consistency check.
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  // Check dimensions of the tensors.
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          std::int64_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                int32_t input_val = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                // Accumulate with 64 bits accumulator.
+                // int64_t += int8_t * int16_t so the highest value we can
+                // get from each accumulation is [-127, 127] * ([-32768,
+                // 32767] -
+                // [-32768, 32767]), which is [-8322945, 8322945].
+                // log2(8322945) = 22.99.
+                acc += filter_val * input_val;
+              }
+            }
+          }
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          scaled_acc = std::max(scaled_acc, output_activation_min);
+          scaled_acc = std::min(scaled_acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int16_t>(scaled_acc);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h

@@ -0,0 +1,289 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+inline void DepthwiseConvPerChannel(
+    const DepthwiseParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data) {
+  // Get parameters.
+  // TODO(b/141565753): Re-introduce ScopedProfilingLabel on Micro.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const int32_t input_offset = params.input_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Check dimensions of the tensors.
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          for (int m = 0; m < depth_multiplier; ++m) {
+            const int output_channel = m + in_channel * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            int32_t acc = 0;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // Zero padding by omitting the areas outside the image.
+                const bool is_point_inside_image =
+                    (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height);
+                if (is_point_inside_image) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, output_channel)];
+                  // Accumulate with 32 bits accumulator.
+                  // In the nudging process during model quantization, we force
+                  // real value of 0.0 be represented by a quantized value. This
+                  // guarantees that the input_offset is a int8_t, even though
+                  // it is represented using int32_t. int32_t += int8_t *
+                  // (int8_t - int8_t) so the highest value we can get from each
+                  // accumulation is [-127, 127] * ([-128, 127] -
+                  // [-128, 127]), which is [-32512, 32512]. log2(32512)
+                  // = 14.98, which means we can accumulate at least 2^16
+                  // multiplications without overflow. The accumulator is
+                  // applied to a filter so the accumulation logic will hold as
+                  // long as the filter size (filter_y * filter_x * in_channel)
+                  // does not exceed 2^16, which is the case in all the models
+                  // we have seen so far.
+                  // TODO(b/174275578): Add a check to make sure the
+                  // accumulator depth is smaller than 2^16.
+                  acc += filter_val * (input_val + input_offset);
+                }
+              }
+            }
+            if (bias_data) {
+              acc += bias_data[output_channel];
+            }
+            acc = MultiplyByQuantizedMultiplier(
+                acc, output_multiplier[output_channel],
+                output_shift[output_channel]);
+            acc += output_offset;
+            acc = std::max(acc, output_activation_min);
+            acc = std::min(acc, output_activation_max);
+            output_data[Offset(output_shape, batch, out_y, out_x,
+                               output_channel)] = static_cast<int8_t>(acc);
+          }
+        }
+      }
+    }
+  }
+}
+
+inline void DepthwiseConvPerChannel(
+    const DepthwiseParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const std::int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  // Get parameters.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Check dimensions of the tensors.
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          for (int m = 0; m < depth_multiplier; ++m) {
+            const int output_channel = m + in_channel * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            std::int64_t acc = 0;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // Zero padding by omitting the areas outside the image.
+                const bool is_point_inside_image =
+                    (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height);
+                if (is_point_inside_image) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, output_channel)];
+                  // Accumulate with 64 bits accumulator.
+                  // We assume maximum of 2^16 accumulations as with the 8-bit
+                  // case so actually the value in the accumulator should not
+                  // exceed 40 bits
+                  acc += static_cast<int64_t>(filter_val) *
+                         static_cast<int64_t>(input_val);
+                }
+              }
+            }
+            if (bias_data) {
+              acc += bias_data[output_channel];
+            }
+            int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+                acc, output_multiplier[output_channel],
+                output_shift[output_channel]);
+            scaled_acc = std::max(scaled_acc, output_activation_min);
+            scaled_acc = std::min(scaled_acc, output_activation_max);
+            output_data[Offset(output_shape, batch, out_y, out_x,
+                               output_channel)] =
+                static_cast<int16_t>(scaled_acc);
+          }
+        }
+      }
+    }
+  }
+}
+
+inline void DepthwiseConvHybridPerChannel(
+    const DepthwiseParams& params, float* scaling_factors_ptr,
+    const RuntimeShape& input_shape, const int8_t* input_data,
+    const RuntimeShape& filter_shape, const int8_t* filter_data,
+    const RuntimeShape& bias_shape, const float* bias_data,
+    const RuntimeShape& output_shape, float* output_data,
+    const float* per_channel_scale, int32_t* input_offset) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  // Check dimensions of the tensors.
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int bias_depth = bias_shape.FlatSize();
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_depth, output_depth);
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          for (int m = 0; m < depth_multiplier; ++m) {
+            const int output_channel = m + in_channel * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            int32_t acc = 0;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // Zero padding by omitting the areas outside the image.
+                const bool is_point_inside_image =
+                    (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height);
+                if (is_point_inside_image) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, output_channel)];
+                  acc += filter_val * (input_val - input_offset[batch]);
+                }
+              }
+            }
+            float acc_float = static_cast<float>(acc);
+            acc_float *=
+                per_channel_scale[output_channel] * scaling_factors_ptr[batch];
+            if (bias_data && output_channel < bias_depth) {
+              acc_float += bias_data[output_channel];
+            }
+            output_data[Offset(output_shape, batch, out_y, out_x,
+                               output_channel)] =
+                ActivationFunctionWithMinMax(acc_float, output_activation_min,
+                                             output_activation_max);
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h

@@ -0,0 +1,109 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data) {
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = output_shape.Dims(0);
+  const int output_depth = output_shape.Dims(1);
+  TFLITE_DCHECK_LE(output_depth, filter_shape.Dims(filter_dim_count - 2));
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      int32_t acc = 0;
+      for (int d = 0; d < accum_depth; ++d) {
+        int32_t input_val = input_data[b * accum_depth + d];
+        int32_t filter_val = filter_data[out_c * accum_depth + d];
+        acc += (filter_val + filter_offset) * (input_val + input_offset);
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc += output_offset;
+      acc = std::max(acc, output_activation_min);
+      acc = std::min(acc, output_activation_max);
+      output_data[out_c + output_depth * b] = static_cast<int8_t>(acc);
+    }
+  }
+}
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = output_shape.Dims(output_dim_count - 1);
+  TFLITE_DCHECK_LE(output_depth, filter_shape.Dims(filter_dim_count - 2));
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      int64_t acc = 0;
+      for (int d = 0; d < accum_depth; ++d) {
+        int32_t input_val = input_data[b * accum_depth + d];
+        int32_t filter_val = filter_data[out_c * accum_depth + d];
+        acc += (filter_val + filter_offset) * input_val;
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+      int32_t acc_scaled =
+          MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc_scaled = std::max(acc_scaled, output_activation_min);
+      acc_scaled = std::min(acc_scaled, output_activation_max);
+      output_data[out_c + output_depth * b] = static_cast<int16_t>(acc_scaled);
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h

@@ -0,0 +1,65 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void L2Normalization(int32_t input_zero_point, int32_t outer_size,
+                            int32_t depth, const int8_t* input_data,
+                            int8_t* output_data) {
+  static constexpr int8_t kMinInt8 = std::numeric_limits<int8_t>::min();
+  static constexpr int8_t kMaxInt8 = std::numeric_limits<int8_t>::max();
+  // The output scale must be in sync with Prepare().
+  // Output is in 1/128 scale so the actual output range is nudged from [-1, 1]
+  // to [-1, 127/128].
+  static constexpr int32_t kOutputScale = 7;
+  for (int outer_index = 0; outer_index < outer_size; ++outer_index) {
+    // int32_t = (int8_t - int8_t) ^ 2.
+    // ([-128, 127] - [-128, 127]) ^ 2 = [0, (2^8 - 1)^2] so the accumulator is
+    // safe from overflowing in at least 2^16 steps.
+    int32_t acc = 0;
+    for (int inner_index = 0; inner_index < depth; ++inner_index) {
+      int32_t input =
+          input_data[depth * outer_index + inner_index] - input_zero_point;
+      acc += input * input;
+    }
+    int32_t inv_l2norm_multiplier;
+    int inv_l2norm_shift;
+    GetInvSqrtQuantizedMultiplierExp(acc, kReverseShift, &inv_l2norm_multiplier,
+                                     &inv_l2norm_shift);
+
+    for (int inner_index = 0; inner_index < depth; ++inner_index) {
+      int32_t input =
+          input_data[depth * outer_index + inner_index] - input_zero_point;
+
+      // Rescale and downcast. Rescale is folded into the division.
+      int32_t output_in_q24 = MultiplyByQuantizedMultiplier(
+          input, inv_l2norm_multiplier, inv_l2norm_shift + kOutputScale);
+      output_in_q24 =
+          std::min(static_cast<int32_t>(kMaxInt8),
+                   std::max(static_cast<int32_t>(kMinInt8), output_in_q24));
+      output_data[depth * outer_index + inner_index] =
+          static_cast<int8_t>(output_in_q24);
+    }
+  }
+}
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h

@@ -0,0 +1,119 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_
+
+#include <limits>
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void Logistic(int32_t input_zero_point, int32_t input_range_radius,
+                     int32_t input_multiplier, int32_t input_left_shift,
+                     int32_t input_size, const int8_t* input_data,
+                     int8_t* output_data) {
+  // Integer bits must be in sync with Prepare() function.
+  static constexpr int32_t kInputIntegerBits = 4;
+  static constexpr int32_t kOutputIntegerBits = 8;
+  static constexpr int8_t kMinInt8 = std::numeric_limits<int8_t>::min();
+  static constexpr int8_t kMaxInt8 = std::numeric_limits<int8_t>::max();
+  static constexpr int32_t kOutputZeroPoint = -128;
+
+  for (int i = 0; i < input_size; ++i) {
+    const int32_t input =
+        static_cast<int32_t>(input_data[i]) - input_zero_point;
+    if (input <= -input_range_radius) {
+      output_data[i] = kMinInt8;
+    } else if (input >= input_range_radius) {
+      output_data[i] = kMaxInt8;
+    } else {
+      const int32_t input_in_q4 = MultiplyByQuantizedMultiplier(
+          input, input_multiplier, input_left_shift);
+      using FixedPoint4 = gemmlowp::FixedPoint<int32_t, kInputIntegerBits>;
+      const int32_t output_in_q0 =
+          gemmlowp::logistic(FixedPoint4::FromRaw(input_in_q4)).raw();
+
+      // Rescale and downcast.
+      using gemmlowp::RoundingDivideByPOT;
+      int32_t output_in_q23 =
+          RoundingDivideByPOT(output_in_q0, 31 - kOutputIntegerBits);
+      output_in_q23 = std::min(std::max(output_in_q23 + kOutputZeroPoint,
+                                        static_cast<int32_t>(kMinInt8)),
+                               static_cast<int32_t>(kMaxInt8));
+      output_data[i] = static_cast<int8_t>(output_in_q23);
+    }
+  }
+}
+
+inline void Logistic(int32_t input_multiplier, int32_t input_left_shift,
+                     int32_t input_size, const int16_t* ptr_input_data,
+                     int16_t* ptr_output_data) {
+  // We use the LUT for sigmoid and take into account, that
+  // tanh(x) = 2*sigmoid(2*x) - 1
+
+  // We scale by 3/4 to expand range [-8,8]->[-10.7,10.7].
+  // In case of general parameter scale, multiplier 3 is taken into account
+  // in TanhPrepare function and it is included in
+  // input_multiplier already.
+
+  TFLITE_DCHECK_GE(input_left_shift, 0);
+  if (input_multiplier == 0) {  // power of two case
+    input_multiplier = 3 << input_left_shift;
+    input_left_shift = 0;
+  }
+
+  int32_t round = (input_left_shift > 0) ? 1 << (input_left_shift - 1) : 0;
+
+  for (int i = 0; i < input_size; ++i, ptr_input_data++, ptr_output_data++) {
+    int32_t input_data =
+        ((*ptr_input_data) * input_multiplier + round) >> input_left_shift;
+
+    // We do interpolation on unsigned values.
+    uint32_t abs_input_data = abs(input_data);
+
+    // We divide by 2 power of 9, because
+    // we need to divide by 2 in power of 7 for
+    // the input conversion + 1/4 from the scale above.
+
+    // Define uh as uint32_t type not to make this function overflow.
+    uint32_t uh = abs_input_data >> 9;
+    uint32_t result;
+
+    if (uh >= 255) {
+      // Saturate to maximum.
+      result = 0x7FFF << 10;
+    } else {
+      uint32_t ua = sigmoid_table_uint16[uh];
+      uint32_t ub = sigmoid_table_uint16[uh + 1];
+      uint32_t ut = abs_input_data & 0x1ff;
+      // Interpolation is done using the fractional bit.
+      result = (ua << 9) + ut * (ub - ua);
+    }
+
+    result = (input_data >= 0) ? (result + (1 << 9))
+                               : ((1 << (16 + 9)) - result + (1 << 9) - 1);
+
+    // Back to 16-bit.
+    result >>= 10;
+
+    *ptr_output_data = result;
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/mean.h

@@ -0,0 +1,77 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+template <typename integer_type>
+inline void Mean(const tflite::MeanParams& op_params, int32_t multiplier,
+                 int32_t shift, const RuntimeShape& unextended_input_shape,
+                 const integer_type* input_data, int32_t input_zero_point,
+                 const RuntimeShape& unextended_output_shape,
+                 integer_type* output_data, int32_t output_zero_point) {
+  // Current implementation only supports dimension equals 4 and simultaneous
+  // reduction over width and height.
+  TFLITE_CHECK_EQ(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_CHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+  const int output_batch = output_shape.Dims(0);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int output_depth = output_shape.Dims(3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int num_elements_in_axis = input_width * input_height;
+
+  TFLITE_CHECK_EQ(op_params.axis_count, 2);
+  TFLITE_CHECK((op_params.axis[0] == 1 && op_params.axis[1] == 2) ||
+               (op_params.axis[0] == 2 && op_params.axis[1] == 1));
+  TFLITE_CHECK_EQ(output_height, 1);
+  TFLITE_CHECK_EQ(output_width, 1);
+
+  static constexpr int32_t kMinInt = std::numeric_limits<integer_type>::min();
+  static constexpr int32_t kMaxInt = std::numeric_limits<integer_type>::max();
+
+  for (int out_b = 0; out_b < output_batch; ++out_b) {
+    for (int out_d = 0; out_d < output_depth; ++out_d) {
+      int32_t acc = 0;
+      for (int in_h = 0; in_h < input_height; ++in_h) {
+        for (int in_w = 0; in_w < input_width; ++in_w) {
+          acc += input_data[Offset(input_shape, out_b, in_h, in_w, out_d)] -
+                 input_zero_point;
+        }
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, multiplier, shift);
+      acc = acc > 0 ? (acc + num_elements_in_axis / 2) / num_elements_in_axis
+                    : (acc - num_elements_in_axis / 2) / num_elements_in_axis;
+      acc += output_zero_point;
+      acc = std::min(std::max(acc, kMinInt), kMaxInt);
+      output_data[Offset(output_shape, out_b, 0, 0, out_d)] =
+          static_cast<integer_type>(acc);
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h

@@ -0,0 +1,131 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_
+
+#include "fixedpoint/fixedpoint.h"
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+template <typename T>
+inline void MulElementwise(int size, const ArithmeticParams& params,
+                           const T* input1_data, const T* input2_data,
+                           T* output_data) {
+  for (int i = 0; i < size; ++i) {
+    const int32_t input1_val = params.input1_offset + input1_data[i];
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t unclamped_result =
+        params.output_offset +
+        MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                      params.output_multiplier,
+                                      params.output_shift);
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, unclamped_result));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+template <typename T>
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const T* input1_data,
+                const RuntimeShape& input2_shape, const T* input2_data,
+                const RuntimeShape& output_shape, T* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  ruy::profiler::ScopeLabel label("Mul/8bit");
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  MulElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+// Mul with 16 bit inputs and int8_t outputs.
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int16_t* input1_data,
+                const RuntimeShape& input2_shape, const int16_t* input2_data,
+                const RuntimeShape& output_shape, int8_t* output_data) {
+  ruy::profiler::ScopeLabel label("Mul/Int16Int8");
+  int32_t output_offset = params.output_offset;
+  int32_t output_activation_min = params.quantized_activation_min;
+  int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    // F0 uses 0 integer bits, range [-1, 1].
+    using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+
+    F0 unclamped_result =
+        F0::FromRaw(input1_data[i]) * F0::FromRaw(input2_data[i]);
+    int16_t rescaled_result =
+        gemmlowp::RoundingDivideByPOT(unclamped_result.raw(), 8);
+    int16_t clamped_result = std::min<int16_t>(
+        output_activation_max - output_offset, rescaled_result);
+    clamped_result = std::max<int16_t>(output_activation_min - output_offset,
+                                       clamped_result);
+    output_data[i] = output_offset + clamped_result;
+  }
+}
+
+template <typename T>
+inline void BroadcastMul4DSlow(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastMul4DSlow");
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  // The input shapes are extended as part of NdArrayDesc initialization.
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              params.input1_offset +
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)];
+          const int32_t input2_val =
+              params.input2_offset +
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+          const int32_t unclamped_result =
+              params.output_offset +
+              MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                            params.output_multiplier,
+                                            params.output_shift);
+          const int32_t clamped_output = std::min(
+              params.quantized_activation_max,
+              std::max(params.quantized_activation_min, unclamped_result));
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              static_cast<T>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h

@@ -0,0 +1,262 @@
+/* Copyright 2018 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_
+
+#include <limits>
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline bool AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const int8_t* input_data,
+                        const RuntimeShape& output_shape, int8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int32_t acc = 0;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              acc +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          if (filter_count == 0) return false;
+          // Round to the closest integer value.
+          acc = acc > 0 ? (acc + filter_count / 2) / filter_count
+                        : (acc - filter_count / 2) / filter_count;
+          acc = std::max(acc, params.quantized_activation_min);
+          acc = std::min(acc, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int8_t>(acc);
+        }
+      }
+    }
+  }
+  return true;
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const int8_t* input_data, const RuntimeShape& output_shape,
+                    int8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_GE(params.quantized_activation_min,
+                   std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_LE(params.quantized_activation_max,
+                   std::numeric_limits<int8_t>::max());
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int8_t max = std::numeric_limits<int8_t>::lowest();
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          max = std::max<int8_t>(max, params.quantized_activation_min);
+          max = std::min<int8_t>(max, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int8_t>(max);
+        }
+      }
+    }
+  }
+}
+
+inline bool AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const int16_t* input_data,
+                        const RuntimeShape& output_shape,
+                        int16_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int32_t acc = 0;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              acc +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          if (filter_count == 0) return false;
+          // Round to the closest integer value.
+          acc = acc > 0 ? (acc + filter_count / 2) / filter_count
+                        : (acc - filter_count / 2) / filter_count;
+          acc = std::max(acc, params.quantized_activation_min);
+          acc = std::min(acc, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int16_t>(acc);
+        }
+      }
+    }
+  }
+  return true;
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const int16_t* input_data, const RuntimeShape& output_shape,
+                    int16_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_GE(params.quantized_activation_min,
+                   std::numeric_limits<int16_t>::min());
+  TFLITE_DCHECK_LE(params.quantized_activation_max,
+                   std::numeric_limits<int16_t>::max());
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int16_t max = std::numeric_limits<int16_t>::lowest();
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          max = std::max<int16_t>(max, params.quantized_activation_min);
+          max = std::min<int16_t>(max, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int16_t>(max);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h

@@ -0,0 +1,116 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_
+
+#include <limits>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void Tanh(int32_t input_zero_point, int32_t input_range_radius,
+                 int32_t input_multiplier, int32_t input_shift,
+                 const RuntimeShape& input_shape, const int8_t* input_data,
+                 const RuntimeShape& output_shape, int8_t* output_data) {
+  // Integer bits must be in sync with Prepare() function.
+  static constexpr int32_t kInputIntegerBits = 4;
+  static constexpr int32_t kOutputScale = 7;
+  static constexpr int32_t kMinInt8 = std::numeric_limits<int8_t>::min();
+  static constexpr int32_t kMaxInt8 = std::numeric_limits<int8_t>::max();
+  using F4 = gemmlowp::FixedPoint<int32_t, kInputIntegerBits>;
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input =
+        static_cast<int32_t>(input_data[i]) - input_zero_point;
+    if (input <= -input_range_radius) {
+      output_data[i] = kMinInt8;
+    } else if (input >= input_range_radius) {
+      output_data[i] = kMaxInt8;
+    } else {
+      const int32_t input_in_q4 =
+          MultiplyByQuantizedMultiplier(input, input_multiplier, input_shift);
+      const int32_t output_in_q0 =
+          gemmlowp::tanh(F4::FromRaw(input_in_q4)).raw();
+
+      // Rescale and downcast.
+      using gemmlowp::RoundingDivideByPOT;
+      int32_t output_in_q24 =
+          RoundingDivideByPOT(output_in_q0, 31 - kOutputScale);
+      output_in_q24 = std::min(std::max(output_in_q24, kMinInt8), kMaxInt8);
+      output_data[i] = static_cast<int8_t>(output_in_q24);
+    }
+  }
+}
+
+inline void Tanh(int32_t input_multiplier, int32_t input_left_shift,
+                 const RuntimeShape& input_shape, const int16_t* ptr_input_data,
+                 const RuntimeShape& output_shape, int16_t* ptr_output_data) {
+  // We use the LUT for sigmoid and take into account, that
+  // tanh(x) = 2*sigmoid(2*x) - 1
+
+  // We scale by 3/4 to expand range [-8,8]->[-10.7,10.7].
+  // In case of general parameter scale, multiplier 3 is taken into account
+  // in TanhPrepare function and it is included in
+  // input_multiplier already.
+
+  if (input_multiplier == 0) {  // power of two case
+    input_multiplier = 3 << input_left_shift;
+    input_left_shift = 0;
+  }
+
+  int32_t round = (input_left_shift > 0) ? 1 << (input_left_shift - 1) : 0;
+
+  int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i, ptr_input_data++, ptr_output_data++) {
+    int32_t input_data =
+        ((*ptr_input_data) * input_multiplier + round) >> input_left_shift;
+
+    uint32_t abs_input_data = abs(input_data);
+    uint32_t uh = abs_input_data >> 8;
+    int32_t result;
+
+    if (uh >= 255) {
+      // Saturate to maximum.
+      result = 0xFFFF << 8;
+    } else {
+      uint32_t ua = sigmoid_table_uint16[uh];
+      uint32_t ub = sigmoid_table_uint16[uh + 1];
+
+      uint8_t ut = abs_input_data & 0xFF;
+
+      result = (ua << 8) + ut * (ub - ua);
+    }
+
+    result = (input_data >= 0)
+                 ? (result - (1 << (14 + 9)) + (1 << (9 - 2)))
+                 : (-result + (1 << (14 + 9)) + (1 << (9 - 2)) - 1);
+
+    // Convert back to 16-bit.
+    result >>= (9 - 1);
+
+    *ptr_output_data = result;
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h

@@ -0,0 +1,221 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+// Fixed-point per-channel-quantization transpose convolution reference kernel.
+inline void TransposeConv(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data, const RuntimeShape& im2col_shape, int8_t* im2col_data,
+    int32_t* scratch_buffer) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int32_t input_offset = params.input_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_activation_min = std::numeric_limits<int8_t>::min();
+  const int32_t output_activation_max = std::numeric_limits<int8_t>::max();
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  const int num_elements = output_shape.FlatSize();
+  // We need to initialize scratch_buffer to all 0s, as we apply the same
+  // 'scatter' based trick as in float version.
+  memset(scratch_buffer, 0, num_elements * sizeof(int32_t));
+
+  // Loop through input elements one at a time.
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int in_y = 0; in_y < input_height; ++in_y) {
+      for (int in_x = 0; in_x < input_width; ++in_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          // Loop through the output elements it will influence.
+          const int out_x_origin = (in_x * stride_width) - pad_width;
+          const int out_y_origin = (in_y * stride_height) - pad_height;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int out_channel = 0; out_channel < output_depth;
+                   ++out_channel) {
+                // Compute output element location.
+                const int out_x = out_x_origin + filter_x;
+                const int out_y = out_y_origin + filter_y;
+                // We cannot accumulate out of bounds.
+                if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) &&
+                    (out_y < output_height)) {
+                  const int8_t input_value = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  const int8_t filter_value =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                        out_channel)] +=
+                      (input_value + input_offset) * filter_value;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          int32_t acc = scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                              out_channel)];
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          acc += output_offset;
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+// int16_t input (zero_point=0), int8_t filter, int64 accumulator
+inline void TransposeConv(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const std::int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data, const RuntimeShape& im2col_shape, int8_t* im2col_data,
+    std::int64_t* scratch_buffer) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int32_t output_activation_min = std::numeric_limits<int16_t>::min();
+  const int32_t output_activation_max = std::numeric_limits<int16_t>::max();
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  const int num_elements = output_shape.FlatSize();
+  // We need to initialize scratch_buffer to all 0s, as we apply the same
+  // 'scatter' based trick as in float version.
+  memset(scratch_buffer, 0, num_elements * sizeof(std::int64_t));
+
+  // Loop through input elements one at a time.
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int in_y = 0; in_y < input_height; ++in_y) {
+      for (int in_x = 0; in_x < input_width; ++in_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          // Loop through the output elements it will influence.
+          const int out_x_origin = (in_x * stride_width) - pad_width;
+          const int out_y_origin = (in_y * stride_height) - pad_height;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int out_channel = 0; out_channel < output_depth;
+                   ++out_channel) {
+                // Compute output element location.
+                const int out_x = out_x_origin + filter_x;
+                const int out_y = out_y_origin + filter_y;
+                // We cannot accumulate out of bounds.
+                if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) &&
+                    (out_y < output_height)) {
+                  const int32_t input_value = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  const int32_t filter_value =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                        out_channel)] +=
+                      input_value * filter_value;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          std::int64_t acc = scratch_buffer[Offset(output_shape, batch, out_y,
+                                                   out_x, out_channel)];
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          scaled_acc = std::max(scaled_acc, output_activation_min);
+          scaled_acc = std::min(scaled_acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int16_t>(scaled_acc);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/l2normalization.h

@@ -0,0 +1,90 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_
+
+#include <algorithm>
+#include <cmath>
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void L2Normalization(const tflite::L2NormalizationParams& op_params,
+                            const RuntimeShape& input_shape,
+                            const float* input_data,
+                            const RuntimeShape& output_shape,
+                            float* output_data, float epsilon = 1e-6) {
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+  for (int i = 0; i < outer_size; ++i) {
+    float squared_l2_norm = 0;
+    for (int c = 0; c < depth; ++c) {
+      const float val = input_data[depth * i + c];
+      squared_l2_norm += val * val;
+    }
+    float l2_norm = std::sqrt(squared_l2_norm);
+    l2_norm = std::max(l2_norm, epsilon);
+    for (int c = 0; c < depth; ++c) {
+      output_data[depth * i + c] = input_data[depth * i + c] / l2_norm;
+    }
+  }
+}
+
+inline void L2Normalization(const tflite::L2NormalizationParams& op_params,
+                            const RuntimeShape& input_shape,
+                            const uint8_t* input_data,
+                            const RuntimeShape& output_shape,
+                            uint8_t* output_data) {
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int32_t input_zero_point = op_params.input_zero_point;
+
+  for (int i = 0; i < outer_size; ++i) {
+    int32_t square_l2_norm = 0;
+    for (int c = 0; c < depth; c++) {
+      int32_t diff = input_data[depth * i + c] - input_zero_point;
+      square_l2_norm += diff * diff;
+    }
+    int32_t inv_l2norm_multiplier;
+    int inv_l2norm_shift;
+    GetInvSqrtQuantizedMultiplierExp(square_l2_norm, kReverseShift,
+                                     &inv_l2norm_multiplier, &inv_l2norm_shift);
+    for (int c = 0; c < depth; c++) {
+      int32_t diff = input_data[depth * i + c] - input_zero_point;
+      int32_t rescaled_diff = MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          128 * diff, inv_l2norm_multiplier, inv_l2norm_shift);
+      int32_t unclamped_output_val = 128 + rescaled_diff;
+      int32_t output_val =
+          std::min(static_cast<int32_t>(255),
+                   std::max(static_cast<int32_t>(0), unclamped_output_val));
+      output_data[depth * i + c] = static_cast<uint8_t>(output_val);
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/leaky_relu.h

@@ -0,0 +1,69 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_LEAKY_RELU_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LEAKY_RELU_H_
+
+#include <algorithm>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void LeakyRelu(const tflite::LeakyReluParams& params,
+                      const RuntimeShape& input_shape, const float* input_data,
+                      const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    const float val = input_data[i];
+    // Note that alpha might be > 1 or < 0, so we don't use std::max here.
+    output_data[i] = val > 0 ? val : val * params.alpha;
+  }
+}
+
+template <typename T>
+inline void QuantizeLeakyRelu(const LeakyReluParams& params,
+                              const RuntimeShape& input_shape,
+                              const T* input_data,
+                              const RuntimeShape& output_shape,
+                              T* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  static const int32_t quantized_min = std::numeric_limits<T>::min();
+  static const int32_t quantized_max = std::numeric_limits<T>::max();
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input_value = input_data[i] - params.input_offset;
+    int32_t unclamped_output;
+    if (input_value >= 0) {
+      unclamped_output = params.output_offset +
+                         MultiplyByQuantizedMultiplier(
+                             input_value, params.output_multiplier_identity,
+                             params.output_shift_identity);
+    } else {
+      unclamped_output = params.output_offset +
+                         MultiplyByQuantizedMultiplier(
+                             input_value, params.output_multiplier_alpha,
+                             params.output_shift_alpha);
+    }
+    const T clamped_output =
+        std::min(quantized_max, std::max(quantized_min, unclamped_output));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LEAKY_RELU_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/log_softmax.h

@@ -0,0 +1,256 @@
+/* Copyright 2021 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_KERNELS_INTERNAL_REFERENCE_LOG_SOFTMAX_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOG_SOFTMAX_H_
+
+#include <algorithm>
+#include <cstddef>
+#include <limits>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void LogSoftmax(const SoftmaxParams& params,
+                       const RuntimeShape& input_shape, const float* input_data,
+                       const RuntimeShape& output_shape, float* output_data) {
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+
+  for (int i = 0; i < outer_size; ++i) {
+    // Find max element value which we'll use to ensure numerical stability
+    // taking advantage of the following equality:
+    // log(exp(x[i])/sum(exp(x[i]))) == log(exp(x[i]+C)/sum(exp(x[i]+C)))
+    float max = std::numeric_limits<float>::lowest();
+    for (int c = 0; c < depth; ++c) {
+      max = std::max(max, input_data[i * depth + c]);
+    }
+
+    // Compute sum.
+    float sum = 0.f;
+    for (int c = 0; c < depth; ++c) {
+      sum += std::exp(input_data[i * depth + c] - max);
+    }
+
+    // Compute result.
+    const float log_sum = std::log(sum);
+    for (int c = 0; c < depth; ++c) {
+      output_data[i * depth + c] = input_data[i * depth + c] - max - log_sum;
+    }
+  }
+}
+
+inline void LogSoftmax(const SoftmaxParams& params,
+                       const RuntimeShape& input_shape,
+                       const uint8_t* input_data,
+                       const RuntimeShape& output_shape, uint8_t* output_data) {
+  const int32_t input_multiplier = params.input_multiplier;
+  const int32_t input_left_shift = params.input_left_shift;
+  const int32_t reverse_scaling_divisor = params.reverse_scaling_divisor;
+  const int32_t reverse_scaling_right_shift =
+      params.reverse_scaling_right_shift;
+  const int diff_min = params.diff_min;
+  // The representation chosen for the input to the exp() function is Q5.26.
+  // We need to leave extra space since values that we skip might be as large
+  // as -32 before multiplying by input_beta_multiplier, and therefore as
+  // large as -16 afterwards.  Note that exp(-8) is definitely not
+  // insignificant to accumulation, but exp(-16) definitely is.
+  static constexpr int kScaledDiffIntegerBits = 5;
+  static constexpr int kAccumulationIntegerBits = 12;
+  static constexpr int kOutputIntegerBits = 4;
+  using FixedPointScaledDiff =
+      gemmlowp::FixedPoint<int32_t, kScaledDiffIntegerBits>;
+  using FixedPointAccum =
+      gemmlowp::FixedPoint<int32_t, kAccumulationIntegerBits>;
+
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+
+  for (int i = 0; i < outer_size; ++i) {
+    uint8_t max_in_row = 0;
+    for (int c = 0; c < depth; ++c) {
+      max_in_row = std::max(max_in_row, input_data[i * depth + c]);
+    }
+
+    FixedPointAccum sum_of_exps = FixedPointAccum::Zero();
+    for (int c = 0; c < depth; ++c) {
+      int32_t input_diff =
+          static_cast<int32_t>(input_data[i * depth + c]) - max_in_row;
+      if (input_diff >= diff_min) {
+        const int32_t input_diff_rescaled =
+            MultiplyByQuantizedMultiplierGreaterThanOne(
+                input_diff, input_multiplier, input_left_shift);
+        const FixedPointScaledDiff scaled_diff_f8 =
+            FixedPointScaledDiff::FromRaw(input_diff_rescaled);
+        sum_of_exps = sum_of_exps + gemmlowp::Rescale<kAccumulationIntegerBits>(
+                                        exp_on_negative_values(scaled_diff_f8));
+      }
+    }
+
+    const int32_t fixed_log_sum_of_exps =
+        log_x_for_x_greater_than_or_equal_to_1<kScaledDiffIntegerBits>(
+            sum_of_exps)
+            .raw();
+
+    // rescaled_diff_min is smallest representable in
+    // Q(kScaledDiffIntegerBits).(31-kScaledDiffIntegerBits) plus the
+    // log-sub-exps that will be subtracted in the loop.
+    //
+    // The thresholds diff_min, etc are negative.
+    const int rescaled_diff_min =
+        fixed_log_sum_of_exps + std::numeric_limits<int32_t>::lowest();
+    const int adjusted_diff_min =
+        std::max(static_cast<int32_t>(
+                     diff_min - 1),  // Note use of > below instead of >= above.
+                 MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                     rescaled_diff_min, reverse_scaling_divisor,
+                     -reverse_scaling_right_shift));
+
+    for (int c = 0; c < depth; ++c) {
+      int32_t input_diff =
+          static_cast<int32_t>(input_data[i * depth + c]) - max_in_row;
+      if (input_diff > adjusted_diff_min) {
+        const int32_t input_diff_rescaled =
+            MultiplyByQuantizedMultiplierGreaterThanOne(
+                input_diff, input_multiplier, input_left_shift);
+        int32_t unsat_output =
+            gemmlowp::RoundingDivideByPOT(
+                (input_diff_rescaled - fixed_log_sum_of_exps),
+                31 - kScaledDiffIntegerBits - kOutputIntegerBits) +
+            255;
+
+        output_data[i * depth + c] = static_cast<uint8_t>(
+            std::max(std::min(unsat_output, static_cast<int32_t>(255)),
+                     static_cast<int32_t>(0)));
+      } else {
+        // Set output to smallest value.
+        output_data[i * depth + c] = 0;
+      }
+    }
+  }
+}
+
+template <typename T>
+inline void LogSoftmaxQuantized(const SoftmaxParams& params,
+                                const size_t outer_size, const size_t depth,
+                                const RuntimeShape& input_shape,
+                                const T* input_data,
+                                const RuntimeShape& output_shape,
+                                T* output_data) {
+  const int32_t input_multiplier = params.input_multiplier;
+  const int32_t input_left_shift = params.input_left_shift;
+  const int32_t reverse_scaling_divisor = params.reverse_scaling_divisor;
+  const int32_t reverse_scaling_right_shift =
+      params.reverse_scaling_right_shift;
+  const int diff_min = params.diff_min;
+
+  static constexpr T kMinT8 = std::numeric_limits<T>::min();
+  static constexpr T kMaxT8 = std::numeric_limits<T>::max();
+  static constexpr int32_t kMinInt32 = std::numeric_limits<int32_t>::min();
+
+  // All IntegerBits must agree with Prepare function.
+  // Input is chosen as Q5.26 so exp(-1 * 2^5 * 2^-1) = exp(-16) is negligible.
+  static constexpr int kInputIntegerBits = 5;
+  static constexpr int kAccumulationIntegerBits = 12;
+  static constexpr int kOutputIntegerBits = 4;
+  using F5 = gemmlowp::FixedPoint<int32_t, kInputIntegerBits>;
+  using F12 = gemmlowp::FixedPoint<int32_t, kAccumulationIntegerBits>;
+
+  for (size_t outer_index = 0; outer_index < outer_size; ++outer_index) {
+    T max_in_row = kMinT8;
+    for (size_t inner_index = 0; inner_index < depth; ++inner_index) {
+      max_in_row =
+          std::max(max_in_row, input_data[outer_index * depth + inner_index]);
+    }
+
+    // Accumulator "sum_of_exps_in_q12" is safe from overflowing in 2^12 steps.
+    F12 sum_of_exps_in_q12 = F12::FromRaw(0);
+    for (size_t inner_index = 0; inner_index < depth; ++inner_index) {
+      int32_t input_diff =
+          static_cast<int32_t>(input_data[outer_index * depth + inner_index]) -
+          max_in_row;
+      if (input_diff >= diff_min) {
+        const int32_t input_diff_in_q5 = MultiplyByQuantizedMultiplier(
+            input_diff, input_multiplier, input_left_shift);
+        sum_of_exps_in_q12 =
+            sum_of_exps_in_q12 +
+            gemmlowp::Rescale<kAccumulationIntegerBits>(
+                exp_on_negative_values(F5::FromRaw(input_diff_in_q5)));
+      }
+    }
+
+    const int32_t log_sum_of_exps_in_q5 =
+        log_x_for_x_greater_than_or_equal_to_1<kInputIntegerBits>(
+            sum_of_exps_in_q12)
+            .raw();
+
+    // Potentially reduced the valid range. shifted_log_sum_of_exps_in_q5 is
+    // smallest representable in Q5.26 plus the log_sum_of_exps.
+    const int32_t shifted_log_sum_of_exps_in_q5 =
+        log_sum_of_exps_in_q5 + kMinInt32;
+    const int32_t adjusted_diff_min =
+        std::max(static_cast<int32_t>(diff_min - 1),
+                 MultiplyByQuantizedMultiplier(shifted_log_sum_of_exps_in_q5,
+                                               reverse_scaling_divisor,
+                                               -reverse_scaling_right_shift));
+
+    for (size_t inner_index = 0; inner_index < depth; ++inner_index) {
+      int32_t input_diff =
+          static_cast<int32_t>(input_data[outer_index * depth + inner_index]) -
+          max_in_row;
+      // Note use of > below instead of >= above.
+      if (input_diff > adjusted_diff_min) {
+        const int32_t input_diff_in_q5 = MultiplyByQuantizedMultiplier(
+            input_diff, input_multiplier, input_left_shift);
+
+        // Rescale and downcast.
+        int32_t output_in_q27 =
+            gemmlowp::RoundingDivideByPOT(
+                (input_diff_in_q5 - log_sum_of_exps_in_q5),
+                31 - kInputIntegerBits - kOutputIntegerBits) +
+            kMaxT8;
+
+        output_in_q27 =
+            std::max(std::min(output_in_q27, static_cast<int32_t>(kMaxT8)),
+                     static_cast<int32_t>(kMinT8));
+        output_data[outer_index * depth + inner_index] =
+            static_cast<T>(output_in_q27);
+      } else {
+        output_data[outer_index * depth + inner_index] = kMinT8;
+      }
+    }
+  }
+}
+
+inline void LogSoftmax(const SoftmaxParams& params, const size_t outer_size,
+                       const size_t depth, const RuntimeShape& input_shape,
+                       const int8_t* input_data,
+                       const RuntimeShape& output_shape, int8_t* output_data) {
+  LogSoftmaxQuantized(params, outer_size, depth, input_shape, input_data,
+                      output_shape, output_data);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOG_SOFTMAX_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/logistic.h

@@ -0,0 +1,132 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_
+
+#include <cmath>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/kernels/op_macros.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void Logistic(const RuntimeShape& input_shape, const float* input_data,
+                     const RuntimeShape& output_shape, float* output_data) {
+  const float cutoff_upper = 16.619047164916992188f;
+  const float cutoff_lower = -9.f;
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  // Rational for using approximation in reference kernel.
+  // 0. This approximation gives enough precision for float.
+  // 1. This works around an issue on an embedded chipset where exp() does not
+  // return correctly as expected - exp(x) should return inf when overflown
+  // not 1.701417   IEEE 754 defines representation for inf.
+  // 2. This will speed up calculation and is matching the behavior in the
+  // optimized kernels. (check the definition of scalar_logistic_op<float>)
+
+  for (int i = 0; i < flat_size; i++) {
+    float val = input_data[i];
+    float result;
+    if (val > cutoff_upper) {
+      result = 1.0f;
+    } else if (val < cutoff_lower) {
+      result = std::exp(val);
+    } else {
+      result = 1.f / (1.f + std::exp(-val));
+    }
+    output_data[i] = result;
+  }
+}
+
+// Convenience version that allows, for example, generated-code calls to be
+// uniform between data types.
+inline void Logistic(const LogisticParams&, const RuntimeShape& input_shape,
+                     const float* input_data, const RuntimeShape& output_shape,
+                     float* output_data) {
+  // Drop params: not needed.
+  Logistic(input_shape, input_data, output_shape, output_data);
+}
+
+inline void Logistic(const LogisticParams& params,
+                     const RuntimeShape& input_shape, const int16_t* input_data,
+                     const RuntimeShape& output_shape, int16_t* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    // F0 uses 0 integer bits, range [-1, 1].
+    // This is the return type of math functions such as tanh, logistic,
+    // whose range is in [-1, 1].
+    using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+    // F3 uses 3 integer bits, range [-8, 8], the input range expected here.
+    using F3 = gemmlowp::FixedPoint<std::int16_t, 3>;
+
+    const F3 input = F3::FromRaw(input_data[i]);
+    F0 output = gemmlowp::logistic(input);
+    output_data[i] = output.raw();
+  }
+}
+
+// Quantized int8_t logistic activation.  Cheats by dequantizing and
+// requantizing around the floating point logistic method.  This implementation
+// is slow on platforms without a floating point unit.
+
+// TODO(b/141211002): Delete this int8_t implementation once we can reuse the
+// approach used in TFLite for int8_t Logistic.
+inline void Logistic(const RuntimeShape& input_shape, const int8_t* input_data,
+                     float input_scale, int input_zero_point,
+                     const RuntimeShape& output_shape, int8_t* output_data,
+                     float output_scale, int output_zero_point) {
+  const float cutoff_upper = 16.619047164916992188f;
+  const float cutoff_lower = -9.f;
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  // Rational for using approximation in reference kernel.
+  // 0. This approximation gives enough precision for float.
+  // 1. This works around an issue on an embedded chipset where exp() does not
+  // return correctly as expected - exp(x) should return inf when overflown
+  // not 1.701417   IEEE 754 defines representation for inf.
+  // 2. This will speed up calculation and is matching the behavior in the
+  // optimized kernels. (check the definition of scalar_logistic_op<float>)
+
+  for (int i = 0; i < flat_size; i++) {
+    // Dequantize.
+    float val =
+        static_cast<float>((input_data[i] - input_zero_point) * input_scale);
+    float result;
+    if (val > cutoff_upper) {
+      result = 1.0f;
+    } else if (val < cutoff_lower) {
+      result = std::exp(val);
+    } else {
+      result = 1.f / (1.f + std::exp(-val));
+    }
+    // Requantize
+    int8_t output =
+        static_cast<int8_t>(result / output_scale + output_zero_point);
+    output_data[i] = output;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/maximum_minimum.h

@@ -0,0 +1,64 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T, typename Op, int N = 5>
+void MaximumMinimumBroadcastSlow(const RuntimeShape& unextended_input1_shape,
+                                 const T* input1_data,
+                                 const RuntimeShape& unextended_input2_shape,
+                                 const T* input2_data,
+                                 const RuntimeShape& unextended_output_shape,
+                                 T* output_data, Op op) {
+  // Uses element-wise calculation if broadcast is not required.
+  if (unextended_input1_shape == unextended_input2_shape) {
+    const int flat_size =
+        MatchingElementsSize(unextended_input1_shape, unextended_input2_shape,
+                             unextended_output_shape);
+    for (int i = 0; i < flat_size; ++i) {
+      output_data[i] = op(input1_data[i], input2_data[i]);
+    }
+  } else {
+    TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), N);
+    TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), N);
+    TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), N);
+
+    NdArrayDesc<N> desc1;
+    NdArrayDesc<N> desc2;
+    NdArrayDesc<N> output_desc;
+    NdArrayDescsForElementwiseBroadcast(
+        unextended_input1_shape, unextended_input2_shape, &desc1, &desc2);
+    CopyDimsToDesc(RuntimeShape::ExtendedShape(N, unextended_output_shape),
+                   &output_desc);
+
+    auto maxmin_func = [&](int indexes[N]) {
+      output_data[SubscriptToIndex(output_desc, indexes)] =
+          op(input1_data[SubscriptToIndex(desc1, indexes)],
+             input2_data[SubscriptToIndex(desc2, indexes)]);
+    };
+    NDOpsHelper<N>(output_desc, maxmin_func);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/mul.h

@@ -0,0 +1,166 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_MUL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Element-wise mul that can often be used for inner loop of broadcast Mul as
+// well as the non-broadcast Mul.
+inline void MulElementwise(int size, const ArithmeticParams& params,
+                           const uint8_t* input1_data,
+                           const uint8_t* input2_data, uint8_t* output_data) {
+  for (int i = 0; i < size; ++i) {
+    const int32_t input1_val = params.input1_offset + input1_data[i];
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t unclamped_result =
+        params.output_offset +
+        MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                      params.output_multiplier,
+                                      params.output_shift);
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, unclamped_result));
+    output_data[i] = static_cast<uint8_t>(clamped_output);
+  }
+}
+
+template <typename T>
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const T* input1_data,
+                const RuntimeShape& input2_shape, const T* input2_data,
+                const RuntimeShape& output_shape, T* output_data) {
+  T output_activation_min;
+  T output_activation_max;
+  GetActivationParams(params, &output_activation_min, &output_activation_max);
+
+  const int flat_size =
+      MatchingExtendedShapeFlatSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] * input2_data[i], output_activation_min,
+        output_activation_max);
+  }
+}
+
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const uint8_t* input1_data,
+                const RuntimeShape& input2_shape, const uint8_t* input2_data,
+                const RuntimeShape& output_shape, uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  const int flat_size =
+      MatchingExtendedShapeFlatSize(input1_shape, input2_shape, output_shape);
+
+  MulElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void BroadcastMul4DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const uint8_t* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const uint8_t* input2_data,
+                               const RuntimeShape& output_shape,
+                               uint8_t* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              params.input1_offset +
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)];
+          const int32_t input2_val =
+              params.input2_offset +
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+          const int32_t unclamped_result =
+              params.output_offset +
+              MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                            params.output_multiplier,
+                                            params.output_shift);
+          const int32_t clamped_output = std::min(
+              params.quantized_activation_max,
+              std::max(params.quantized_activation_min, unclamped_result));
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              static_cast<uint8_t>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+void BroadcastMul4DSlow(const ArithmeticParams& params,
+                        const RuntimeShape& unextended_input1_shape,
+                        const T* input1_data,
+                        const RuntimeShape& unextended_input2_shape,
+                        const T* input2_data,
+                        const RuntimeShape& unextended_output_shape,
+                        T* output_data) {
+  T output_activation_min;
+  T output_activation_max;
+  GetActivationParams(params, &output_activation_min, &output_activation_max);
+
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
+                                      unextended_input2_shape, &desc1, &desc2);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < output_shape.Dims(0); ++b) {
+    for (int y = 0; y < output_shape.Dims(1); ++y) {
+      for (int x = 0; x < output_shape.Dims(2); ++x) {
+        for (int c = 0; c < output_shape.Dims(3); ++c) {
+          output_data[Offset(output_shape, b, y, x, c)] =
+              ActivationFunctionWithMinMax(
+                  input1_data[SubscriptToIndex(desc1, b, y, x, c)] *
+                      input2_data[SubscriptToIndex(desc2, b, y, x, c)],
+                  output_activation_min, output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/neg.h

@@ -0,0 +1,37 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_NEG_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void Negate(const RuntimeShape& input_shape, const T* input_data,
+                   const RuntimeShape& output_shape, T* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = -input_data[i];
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/pad.h

@@ -0,0 +1,169 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_PAD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PAD_H_
+
+#include <vector>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// TFLite Pad supports activation tensors with up to 5 dimensions.
+constexpr int PadKernelMaxDimensionCount() { return 5; }
+
+// There are two versions of pad: Pad and PadV2.  In PadV2 there is a second
+// scalar input that provides the padding value.  Therefore pad_value_ptr can be
+// equivalent to a simple input1_data.  For Pad, it should point to a zero
+// value.
+//
+// Note that two typenames are required, so that T=P=int32_t is considered a
+// specialization distinct from P=int32_t.
+template <typename T, typename P>
+inline void PadImpl(const tflite::PadParams& op_params,
+                    const RuntimeShape& input_shape, const T* input_data,
+                    const P* pad_value_ptr, const RuntimeShape& output_shape,
+                    T* output_data) {
+  const RuntimeShape ext_input_shape =
+      RuntimeShape::ExtendedShape(PadKernelMaxDimensionCount(), input_shape);
+  const RuntimeShape ext_output_shape =
+      RuntimeShape::ExtendedShape(PadKernelMaxDimensionCount(), output_shape);
+  TFLITE_DCHECK_LE(op_params.left_padding_count, PadKernelMaxDimensionCount());
+  TFLITE_DCHECK_LE(op_params.right_padding_count, PadKernelMaxDimensionCount());
+
+  // Runtime calls are currently fixed at 5 dimensions. Copy inputs so we can
+  // pad them to 5 dims (yes, we are "padding the padding").
+  int left_padding_copy[PadKernelMaxDimensionCount()];
+  for (int i = 0; i < PadKernelMaxDimensionCount(); i++) {
+    left_padding_copy[i] = 0;
+  }
+  for (int i = 0; i < op_params.left_padding_count; ++i) {
+    left_padding_copy[i + PadKernelMaxDimensionCount() -
+                      op_params.left_padding_count] = op_params.left_padding[i];
+  }
+  int right_padding_copy[PadKernelMaxDimensionCount()];
+  for (int i = 0; i < PadKernelMaxDimensionCount(); i++) {
+    right_padding_copy[i] = 0;
+  }
+  for (int i = 0; i < op_params.right_padding_count; ++i) {
+    right_padding_copy[i + PadKernelMaxDimensionCount() -
+                       op_params.right_padding_count] =
+        op_params.right_padding[i];
+  }
+
+  const int output_batch = ext_output_shape.Dims(0);
+  const int output_plane = ext_output_shape.Dims(1);
+  const int output_height = ext_output_shape.Dims(2);
+  const int output_width = ext_output_shape.Dims(3);
+  const int output_depth = ext_output_shape.Dims(4);
+
+  const int left_b_padding = left_padding_copy[0];
+  const int left_p_padding = left_padding_copy[1];
+  const int left_h_padding = left_padding_copy[2];
+  const int left_w_padding = left_padding_copy[3];
+  const int left_d_padding = left_padding_copy[4];
+
+  const int right_b_padding = right_padding_copy[0];
+  const int right_p_padding = right_padding_copy[1];
+  const int right_h_padding = right_padding_copy[2];
+  const int right_w_padding = right_padding_copy[3];
+  const int right_d_padding = right_padding_copy[4];
+
+  const T pad_value = *pad_value_ptr;
+
+  const T* in_ptr = input_data;
+  T* out_ptr = output_data;
+  for (int out_b = 0; out_b < output_batch; ++out_b) {
+    for (int out_p = 0; out_p < output_plane; ++out_p) {
+      for (int out_h = 0; out_h < output_height; ++out_h) {
+        for (int out_w = 0; out_w < output_width; ++out_w) {
+          for (int out_d = 0; out_d < output_depth; ++out_d) {
+            if (out_b < left_b_padding ||
+                out_b >= output_batch - right_b_padding ||
+                out_p < left_p_padding ||
+                out_p >= output_plane - right_p_padding ||
+                out_h < left_h_padding ||
+                out_h >= output_height - right_h_padding ||
+                out_w < left_w_padding ||
+                out_w >= output_width - right_w_padding ||
+                out_d < left_d_padding ||
+                out_d >= output_depth - right_d_padding) {
+              *out_ptr++ = pad_value;
+            } else {
+              *out_ptr++ = *in_ptr++;
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+template <typename T, typename P>
+inline void Pad(const tflite::PadParams& op_params,
+                const RuntimeShape& input_shape, const T* input_data,
+                const P* pad_value_ptr, const RuntimeShape& output_shape,
+                T* output_data) {
+  PadImpl(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+          output_data);
+}
+
+// The second (pad-value) input can be int32_t when, say, the first is uint8_t.
+template <typename T>
+inline void Pad(const tflite::PadParams& op_params,
+                const RuntimeShape& input_shape, const T* input_data,
+                const int32_t* pad_value_ptr, const RuntimeShape& output_shape,
+                T* output_data) {
+  const T converted_pad_value = static_cast<T>(*pad_value_ptr);
+  PadImpl(op_params, input_shape, input_data, &converted_pad_value,
+          output_shape, output_data);
+}
+
+// This version avoids conflicting template matching.
+template <>
+inline void Pad(const tflite::PadParams& op_params,
+                const RuntimeShape& input_shape, const int32_t* input_data,
+                const int32_t* pad_value_ptr, const RuntimeShape& output_shape,
+                int32_t* output_data) {
+  PadImpl(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+          output_data);
+}
+
+template <typename T, typename P>
+inline void PadImageStyle(const tflite::PadParams& op_params,
+                          const RuntimeShape& input_shape, const T* input_data,
+                          const P* pad_value_ptr,
+                          const RuntimeShape& output_shape, T* output_data) {
+  Pad(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+      output_data);
+}
+
+template <typename P>
+inline void PadImageStyle(const tflite::PadParams& op_params,
+                          const RuntimeShape& input_shape,
+                          const float* input_data, const P* pad_value_ptr,
+                          const RuntimeShape& output_shape,
+                          float* output_data) {
+  Pad(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+      output_data);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PAD_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/pooling.h

@@ -0,0 +1,301 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_POOLING_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline bool AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const float* input_data,
+                        const RuntimeShape& output_shape, float* output_data) {
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          float total = 0.f;
+          float filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              total +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          if (filter_count == 0) return false;
+          const float average = total / filter_count;
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              ActivationFunctionWithMinMax(average, params.float_activation_min,
+                                           params.float_activation_max);
+        }
+      }
+    }
+  }
+  return true;
+}
+
+inline bool AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const uint8_t* input_data,
+                        const RuntimeShape& output_shape,
+                        uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int32_t acc = 0;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              acc +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          if (filter_count == 0) return false;
+          acc = (acc + filter_count / 2) / filter_count;
+          acc = std::max(acc, params.quantized_activation_min);
+          acc = std::min(acc, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<uint8_t>(acc);
+        }
+      }
+    }
+  }
+  return true;
+}
+
+inline void L2Pool(const PoolParams& params, const RuntimeShape& input_shape,
+                   const float* input_data, const RuntimeShape& output_shape,
+                   float* output_data) {
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          float sum_squares = 0.f;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              const float val =
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              sum_squares += val * val;
+              filter_count++;
+            }
+          }
+          const float l2pool_result = std::sqrt(sum_squares / filter_count);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              ActivationFunctionWithMinMax(l2pool_result,
+                                           params.float_activation_min,
+                                           params.float_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const float* input_data, const RuntimeShape& output_shape,
+                    float* output_data) {
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          float max = std::numeric_limits<float>::lowest();
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              ActivationFunctionWithMinMax(max, params.float_activation_min,
+                                           params.float_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const uint8_t* input_data, const RuntimeShape& output_shape,
+                    uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_GE(params.quantized_activation_min, 0);
+  TFLITE_DCHECK_LE(params.quantized_activation_max, 255);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          uint8_t max = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          max = std::max<uint8_t>(max, params.quantized_activation_min);
+          max = std::min<uint8_t>(max, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<uint8_t>(max);
+        }
+      }
+    }
+  }
+}
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc

@@ -0,0 +1,774 @@
+/* Copyright 2019 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 <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <utility>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h"
+
+#if defined(_MSC_VER)
+#define __restrict__ __restrict
+#endif
+
+namespace tflite {
+namespace tensor_utils {
+
+namespace {
+const int32_t kInt16Max = std::numeric_limits<int16_t>::max();
+const int32_t kInt16Min = std::numeric_limits<int16_t>::min();
+}  // namespace
+
+void PortableSymmetricQuantizeFloats(const float* values, const int size,
+                                     int8_t* quantized_values, float* min_value,
+                                     float* max_value, float* scaling_factor) {
+  auto minmax = std::minmax_element(values, values + size);
+  *min_value = *minmax.first;
+  *max_value = *minmax.second;
+
+  PortableSymmetricQuantizeFloats(values, size, quantized_values, *min_value,
+                                  *max_value, scaling_factor);
+}
+
+void PortableSymmetricQuantizeFloats(const float* values, const int size,
+                                     int8_t* quantized_values, float min_value,
+                                     float max_value, float* scaling_factor) {
+  const int32_t kScale = 127;
+  const float range = std::max(std::abs(min_value), std::abs(max_value));
+  if (range == 0) {
+    memset(quantized_values, 0, size * sizeof(int8_t));
+    *scaling_factor = 1;
+    return;
+  }
+  *scaling_factor = range / kScale;
+  const float scaling_factor_inv = kScale / range;
+  for (int i = 0; i < size; ++i) {
+    const int32_t quantized_value =
+        static_cast<int32_t>(TfLiteRound(values[i] * scaling_factor_inv));
+    // Clamp: just in case some odd numeric offset.
+    quantized_values[i] = static_cast<int8_t>(
+        std::min(kScale, std::max(-kScale, quantized_value)));
+  }
+}
+
+void PortableAsymmetricQuantizeFloats(const float* values, const int size,
+                                      int8_t* quantized_values,
+                                      float* scaling_factor, int32_t* offset) {
+  const int32_t kMinScale = -128;
+  const int32_t kMaxScale = 127;
+  const double qmin_double = kMinScale;
+  const double qmax_double = kMaxScale;
+  const auto minmax = std::minmax_element(values, values + size);
+  const double rmin = static_cast<double>(std::min(0.0f, *minmax.first));
+  const double rmax = static_cast<double>(std::max(0.0f, *minmax.second));
+  if (rmin == rmax) {
+    memset(quantized_values, 0, size * sizeof(int8_t));
+    *scaling_factor = 1;
+    *offset = 0;
+    return;
+  } else {
+    double scale = (rmax - rmin) / (qmax_double - qmin_double);
+    const double zero_point_from_min = qmin_double - rmin / scale;
+    const double zero_point_from_max = qmax_double - rmax / scale;
+    const double zero_point_from_min_error =
+        std::abs(qmin_double) + std::abs(rmin / scale);
+    const double zero_point_from_max_error =
+        std::abs(qmax_double) + std::abs(rmax / scale);
+    const double zero_point_double =
+        zero_point_from_min_error < zero_point_from_max_error
+            ? zero_point_from_min
+            : zero_point_from_max;
+    int8_t nudged_zero_point = 0;
+    if (zero_point_double <= qmin_double) {
+      nudged_zero_point = kMinScale;
+    } else if (zero_point_double >= qmax_double) {
+      nudged_zero_point = kMaxScale;
+    } else {
+      nudged_zero_point = static_cast<int8_t>(round(zero_point_double));
+    }
+    *scaling_factor = scale;
+    *offset = nudged_zero_point;
+  }
+  const float scaling_factor_inv = 1.0f / *scaling_factor;
+  for (int i = 0; i < size; ++i) {
+    const int32_t quantized_value = static_cast<int32_t>(
+        TfLiteRound(*offset + values[i] * scaling_factor_inv));
+    quantized_values[i] =
+        std::min(kMaxScale, std::max(kMinScale, quantized_value));
+  }
+}
+
+void PortableMatrixBatchVectorMultiplyAccumulate(const float* matrix,
+                                                 int m_rows, int m_cols,
+                                                 const float* vector,
+                                                 int n_batch, float* result) {
+  float* result_in_batch = result;
+  for (int b = 0; b < n_batch; b++) {
+    const float* matrix_ptr = matrix;
+    for (int r = 0; r < m_rows; r++) {
+      float dot_prod = 0.0f;
+      const float* vector_in_batch = vector + b * m_cols;
+      for (int c = 0; c < m_cols; c++) {
+        dot_prod += *matrix_ptr++ * *vector_in_batch++;
+      }
+      *result_in_batch += dot_prod;
+      ++result_in_batch;
+    }
+  }
+}
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vectors, const float* scaling_factors,
+    int n_batch, float* __restrict__ result) {
+  for (int batch = 0; batch < n_batch; ++batch, vectors += m_cols) {
+    const float batch_scaling_factor = scaling_factors[batch];
+    // Get the address of the first row.
+    const int8_t* row_ptr = matrix;
+    for (int row = 0; row < m_rows; ++row) {
+      // Initialize the dot product sum for the row to 0.
+      int32_t dotprod = 0;
+#if defined(__GNUC__)
+      // Prefetch the row to cache.
+      __builtin_prefetch(row_ptr, 0 /* prefetch for read */,
+                         3 /* temporal locality */);
+#endif
+      for (int col = 0; col < m_cols; ++col, ++row_ptr) {
+        dotprod += (*row_ptr) * (vectors[col]);
+      }  // for col
+      *result += dotprod * batch_scaling_factor;
+      ++result;
+    }  // for row
+  }    // for batch
+}
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vectors, const float* scaling_factors,
+    int n_batch, float* __restrict__ result, const float* per_channel_scale,
+    const int32_t* input_offset, int32_t* scratch, int32_t* row_sums,
+    bool* compute_row_sums, CpuBackendContext* context) {
+  if (input_offset == nullptr) {
+    PortableMatrixBatchVectorMultiplyAccumulate(
+        matrix, m_rows, m_cols, vectors, scaling_factors, n_batch, result);
+    return;
+  }
+  if (!compute_row_sums || *compute_row_sums) {
+    PortableReductionSumVector(matrix, row_sums, m_rows, m_cols);
+    if (compute_row_sums) {
+      *compute_row_sums = false;
+    }
+  }
+
+  for (int batch = 0; batch < n_batch; ++batch, vectors += m_cols) {
+    const float batch_scaling_factor = scaling_factors[batch];
+    const int32_t batch_offset = input_offset[batch];
+    const int8_t* row_ptr = matrix;
+    for (int row = 0; row < m_rows; ++row) {
+      int32_t dotprod = 0;
+      float scale = batch_scaling_factor;
+      if (per_channel_scale) {
+        scale *= per_channel_scale[row];
+      }
+#if defined(__GNUC__)
+      // Prefetch the row to cache.
+      __builtin_prefetch(row_ptr, 0 /* prefetch for read */,
+                         3 /* temporal locality */);
+#endif
+      for (int col = 0; col < m_cols; ++col, ++row_ptr) {
+        dotprod += (*row_ptr) * vectors[col];
+      }  // for col
+      dotprod -= row_sums[row] * batch_offset;
+      *result += dotprod * scale;
+      ++result;
+    }  // for row
+  }    // for batch
+}
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate1x4(
+    const float* __restrict__ matrix, const int32_t* __restrict__ segments,
+    const int32_t* __restrict__ indices, int m_rows, int m_cols,
+    const float* __restrict__ vector, int n_batch, float* __restrict__ result) {
+  const int kBlockSize = 4;
+  TFLITE_DCHECK_EQ(m_cols % kBlockSize, 0);
+  for (int batch = 0; batch < n_batch; batch++) {
+    const float* matrix_ptr = matrix;
+    for (int row = 0; row < m_rows; row++) {
+      float dot_prod = 0.0f;
+      const float* vector_in_batch = vector + batch * m_cols;
+      for (int i = segments[row]; i < segments[row + 1]; i++) {
+        const int block_start_index = indices[i] * kBlockSize;
+        const float* vector_block_in_batch_ptr =
+            vector_in_batch + block_start_index;
+        for (int c = 0; c < kBlockSize; c++) {
+          dot_prod += *matrix_ptr++ * *vector_block_in_batch_ptr++;
+        }
+      }
+      result[batch * m_rows + row] += dot_prod;
+    }
+  }
+}
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate(
+    const float* __restrict__ matrix, const uint8_t* __restrict__ ledger,
+    int m_rows, int m_cols, const float* __restrict__ vector, int n_batch,
+    float* __restrict__ result) {
+  const int kBlockSize = 16;
+  TFLITE_DCHECK_EQ(  // NOLINT
+      m_cols % kBlockSize, 0);
+  for (int batch = 0; batch < n_batch; batch++) {
+    const float* matrix_ptr = matrix;
+    const uint8_t* ledger_ptr = ledger;
+    for (int row = 0; row < m_rows; row++) {
+      float dot_prod = 0.0f;
+      int num_nonzero_blocks = *ledger_ptr++;
+      if (num_nonzero_blocks > 0) {
+        const float* vector_in_batch = vector + batch * m_cols;
+        for (int i = 0; i < num_nonzero_blocks; i++) {
+          const int block_start_index = *ledger_ptr++ * kBlockSize;
+          const float* vector_block_in_batch_ptr =
+              vector_in_batch + block_start_index;
+          for (int c = 0; c < kBlockSize; c++) {
+            dot_prod += *matrix_ptr++ * *vector_block_in_batch_ptr++;
+          }
+        }
+      }
+      result[batch * m_rows + row] += dot_prod;
+    }
+  }
+}
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const uint8_t* ledger, const int m_rows,
+    const int m_cols, const int8_t* __restrict__ vectors,
+    const float* scaling_factors, int n_batch, float* __restrict__ result) {
+  static const int kBlockSize = 16;
+  TFLITE_DCHECK_EQ(  // NOLINT
+      m_cols % kBlockSize, 0);
+  for (int batch = 0; batch < n_batch; ++batch, vectors += m_cols) {
+    const float batch_scaling_factor = scaling_factors[batch];
+    const uint8_t* ledger_ptr = ledger;
+    // Get the address of the first row.
+    const int8_t* row_ptr = matrix;
+    for (int row = 0; row < m_rows; ++row) {
+      // Initialize the dot product sum for the row to 0.
+      int32_t dotprod = 0;
+#if defined(__GNUC__)
+      // Prefetch the row to cache.
+      __builtin_prefetch(row_ptr, 0 /* prefetch for read */,
+                         3 /* temporal locality */);
+#endif
+      int num_nonzero_blocks = *ledger_ptr++;
+      for (int i = 0; i < num_nonzero_blocks; i++) {
+        const int block_start_index = *ledger_ptr++ * kBlockSize;
+        const int8_t* vector_block_ptr = vectors + block_start_index;
+        for (int c = 0; c < kBlockSize; c++) {
+          dotprod += (*row_ptr++) * (*vector_block_ptr++);
+        }  // for block
+      }    // for num_nonzero_blocks
+      result[batch * m_rows + row] += dotprod * batch_scaling_factor;
+    }  // for row
+  }    // for batch
+}
+
+template <typename T>
+void PortableMatrixBatchVectorMultiplyAccumulateImpl(
+    const int8_t* input, const int32_t* bias,
+    const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
+    int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
+    T* output) {
+  const int16_t output_max = std::numeric_limits<T>::max();
+  const int16_t output_min = std::numeric_limits<T>::min();
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int row = 0; row < n_output; ++row) {
+      int32_t acc = bias[row];
+      for (int col = 0; col < n_input; ++col) {
+        int8_t input_val = input[batch * n_input + col];
+        int8_t weights_val = input_to_gate_weights[row * n_input + col];
+        acc += input_val * weights_val;
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, multiplier, shift);
+      acc += output_zp;
+      acc += output[batch * n_output + row];
+      if (acc > output_max) {
+        acc = output_max;
+      }
+      if (acc < output_min) {
+        acc = output_min;
+      }
+      output[batch * n_output + row] = static_cast<T>(acc);
+    }
+  }
+}
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* input, const int32_t* bias,
+    const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
+    int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
+    int32_t* scratch, int16_t* output, CpuBackendContext* context) {
+  PortableMatrixBatchVectorMultiplyAccumulateImpl(
+      input, bias, input_to_gate_weights, multiplier, shift, n_batch, n_input,
+      n_output, output_zp, output);
+}
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* input, const int32_t* bias,
+    const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
+    int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
+    int32_t* scratch, int8_t* output, CpuBackendContext* context) {
+  PortableMatrixBatchVectorMultiplyAccumulateImpl(
+      input, bias, input_to_gate_weights, multiplier, shift, n_batch, n_input,
+      n_output, output_zp, output);
+}
+
+void PortableMatrixBatchVectorMultiply(const int8_t* input,
+                                       int32_t input_zeropoint,
+                                       const int8_t* input_to_gate_weights,
+                                       int32_t input_to_gate_effective_scale_a,
+                                       int32_t input_to_gate_effective_scale_b,
+                                       int32_t n_batch, int32_t n_input,
+                                       int32_t n_cell, int8_t* gate_output,
+                                       int8_t gate_output_zp) {
+  const int32_t int8_max = std::numeric_limits<int8_t>::max();
+  const int32_t int8_min = std::numeric_limits<int8_t>::min();
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int row = 0; row < n_cell; ++row) {
+      int32_t acc = 0;
+      for (int col = 0; col < n_input; ++col) {
+        int32_t input_val = input[batch * n_input + col];
+        int8_t weights_val = input_to_gate_weights[row * n_input + col];
+        acc += (input_val - input_zeropoint) * weights_val;
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, input_to_gate_effective_scale_a,
+                                          input_to_gate_effective_scale_b);
+      acc += gate_output_zp;
+      if (acc > int8_max) {
+        acc = int8_max;
+      }
+      if (acc < int8_min) {
+        acc = int8_min;
+      }
+      gate_output[batch * n_cell + row] = static_cast<int8_t>(acc);
+    }
+  }
+}
+
+void PortableMatrixBatchVectorMultiply(
+    const int16_t* hidden, const int8_t* hidden_to_output_weights,
+    int32_t proj_effective_scale_a, int32_t proj_effective_scale_b,
+    const int32_t* gate_bias, int32_t n_batch, int32_t n_hidden,
+    int32_t n_output, int32_t output_zp, int8_t* proj_output) {
+  const int16_t int8_max = std::numeric_limits<int8_t>::max();
+  const int16_t int8_min = std::numeric_limits<int8_t>::min();
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int row = 0; row < n_output; ++row) {
+      int64_t acc = gate_bias[row];
+      for (int col = 0; col < n_hidden; ++col) {
+        int16_t input_val = hidden[batch * n_hidden + col];
+        int8_t weights_val = hidden_to_output_weights[row * n_hidden + col];
+        int64_t curr = acc;
+        acc += input_val * weights_val;
+        if (input_val * weights_val > 0 && acc < curr) {
+          acc = std::numeric_limits<int32_t>::max();
+        }
+        if (input_val * weights_val < 0 && acc > curr) {
+          acc = std::numeric_limits<int32_t>::min();
+        }
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, proj_effective_scale_a,
+                                          proj_effective_scale_b);
+      acc += output_zp;
+      if (acc > int8_max) {
+        acc = int8_max;
+      }
+      if (acc < int8_min) {
+        acc = int8_min;
+      }
+      proj_output[batch * n_output + row] = acc;
+    }
+  }
+}
+
+void PortableApplyLayerNorm(const int16_t* input,
+                            const int16_t* layer_norm_weights,
+                            const int32_t* bias, int32_t layer_norm_scale_a,
+                            int32_t layer_norm_scale_b, int32_t variance_limit,
+                            int n_batch, int n_input, int16_t* output) {
+  // The square of std::pow(2, 10), which is the extra factor that makes sure
+  // normalized values has enough resolution.
+  static const int kTwoToPower20 = 1 << 20;
+  for (int i = 0; i < n_batch; ++i) {
+    int64_t sum = 0;
+    int64_t sum_sq = 0;
+    for (int j = 0; j < n_input; ++j) {
+      const int32_t index = i * n_input + j;
+      int32_t val = static_cast<int32_t>(input[index]);
+      sum += val;
+      sum_sq += val * val;
+    }
+    int32_t mean =
+        static_cast<int32_t>(static_cast<int64_t>(sum) * 1024 / n_input);
+    // TODO(b/173994730): Avoids overflow but only works for POT n_input.
+    int32_t temp = kTwoToPower20 / n_input;
+    int64_t variance =
+        sum_sq * temp - static_cast<int64_t>(mean) * static_cast<int64_t>(mean);
+    int32_t variance2 = static_cast<int32_t>(variance / kTwoToPower20);
+    if (variance2 < 1) {
+      variance2 = variance_limit;
+    }
+    int32_t stddev_inverse_a;
+    int stddev_inverse_b;
+    GetInvSqrtQuantizedMultiplierExp(variance2, /*reverse_shift*/ -1,
+                                     &stddev_inverse_a, &stddev_inverse_b);
+
+    for (int j = 0; j < n_input; ++j) {
+      const int32_t index = i * n_input + j;
+      int32_t val = static_cast<int32_t>(input[index]);
+      int32_t shifted = 1024 * val - mean;
+      int32_t rescaled = MultiplyByQuantizedMultiplier(
+          shifted, stddev_inverse_a, stddev_inverse_b);
+      // TODO(jianlijianli): Saturate this.
+      int64_t val3 = rescaled * layer_norm_weights[j] + bias[j];
+      int32_t val4 =
+          static_cast<int32_t>((val3 > 0 ? val3 + 512 : val3 - 512) / 1024);
+      int32_t val5 = MultiplyByQuantizedMultiplier(val4, layer_norm_scale_a,
+                                                   layer_norm_scale_b + 12);
+      val5 = std::min(std::max(kInt16Min, val5), kInt16Max);
+      output[index] = static_cast<int16_t>(val5);
+    }
+  }
+}
+
+void PortableApplyLayerNormFloat(const int16_t* input,
+                                 const int16_t* layer_norm_weights,
+                                 int32_t layer_norm_scale_a,
+                                 int32_t layer_norm_scale_b,
+                                 const int32_t* bias, int n_batch, int n_input,
+                                 int16_t* output) {
+  const int32_t int16_max = std::numeric_limits<int16_t>::max();
+  const int32_t int16_min = std::numeric_limits<int16_t>::min();
+  const float layer_norm_scale =
+      layer_norm_scale_a *
+      std::pow(2.0, static_cast<double>(layer_norm_scale_b - 31));
+  const float bias_scale =
+      static_cast<float>(std::pow(2.0, -10)) * layer_norm_scale;
+
+  for (int batch = 0; batch < n_batch; ++batch) {
+    float sum = 0.0f;
+    float sum_sq = 0.0f;
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      const float value = static_cast<float>(input[index]);
+      sum += value;
+      sum_sq += value * value;
+    }
+    const float mean = sum / n_input;
+    float stddev_inv = 0.0f;
+    const float variance = sum_sq / n_input - mean * mean;
+    if (variance == 0) {
+      stddev_inv = 1.0f / std::sqrt(1e-8f);
+    } else {
+      stddev_inv = 1.0f / std::sqrt(variance);
+    }
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      const float normalized_value =
+          (static_cast<float>(input[index]) - mean) * stddev_inv;
+      const float weighted_normalized_value =
+          normalized_value * layer_norm_weights[i] * layer_norm_scale +
+          bias[i] * bias_scale;
+      const int32_t quant_output = static_cast<int32_t>(round(
+          weighted_normalized_value * static_cast<float>(std::pow(2, 12))));
+      output[index] = std::min(int16_max, std::max(int16_min, quant_output));
+    }
+  }
+}
+
+void PortableMatrixScalarMultiplyAccumulate(const int8_t* matrix,
+                                            int32_t scalar, int32_t n_row,
+                                            int32_t n_col, int32_t* output) {
+  for (int i = 0; i < n_row; ++i) {
+    int32_t row_sum = 0;
+    for (int j = 0; j < n_col; ++j) {
+      row_sum += *matrix++;
+    }
+    output[i] += row_sum * scalar;
+  }
+}
+
+void PortableApplySigmoid(const int16_t* input, int32_t n_batch,
+                          int32_t n_input, int16_t* output) {
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int c = 0; c < n_input; c++) {
+      using F3 = gemmlowp::FixedPoint<std::int16_t, 3>;
+      using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+      const int index = batch * n_input + c;
+      F3 sigmoid_input = F3::FromRaw(input[index]);
+      F0 sigmoid_output = gemmlowp::logistic(sigmoid_input);
+      output[index] = sigmoid_output.raw();
+    }
+  }
+}
+
+void PortableApplySigmoidFloat(const int16_t* input, int32_t n_batch,
+                               int32_t n_input, int16_t* output) {
+  const int32_t int16_max = std::numeric_limits<int16_t>::max();
+  const int32_t int16_min = std::numeric_limits<int16_t>::min();
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      const float float_input =
+          input[index] * static_cast<float>(std::pow(2, -12));
+      const float float_output = 1.0f / (1.0f + std::exp(-float_input));
+      const int32_t quant_output = static_cast<int32_t>(
+          float_output * static_cast<float>(std::pow(2, 15)));
+      const int32_t quant_output_clamped =
+          std::min(int16_max, std::max(int16_min, quant_output));
+      output[index] = static_cast<int16_t>(quant_output_clamped);
+    }
+  }
+}
+
+template <int IntegerBits>
+void PortableApplyTanhImpl(const int16_t* input, int32_t n_batch,
+                           int32_t n_input, int16_t* output) {
+  using FX = gemmlowp::FixedPoint<std::int16_t, IntegerBits>;
+  using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      FX tanh_input = FX::FromRaw(input[index]);
+      F0 tanh_output = gemmlowp::tanh(tanh_input);
+      output[index] = tanh_output.raw();
+    }
+  }
+}
+
+void PortableApplyTanh(int32_t integer_bits, const int16_t* input,
+                       int32_t n_batch, int32_t n_input, int16_t* output) {
+  assert(integer_bits <= 6);
+#define DISPATCH_TANH(i)                                       \
+  case i:                                                      \
+    PortableApplyTanhImpl<i>(input, n_batch, n_input, output); \
+    break;
+  switch (integer_bits) {
+    DISPATCH_TANH(0);
+    DISPATCH_TANH(1);
+    DISPATCH_TANH(2);
+    DISPATCH_TANH(3);
+    DISPATCH_TANH(4);
+    DISPATCH_TANH(5);
+    DISPATCH_TANH(6);
+    default:
+      return;
+  }
+#undef DISPATCH_TANH
+}
+
+void PortableApplyTanhFloat(const int16_t* input, int32_t n_batch,
+                            int32_t n_input, int32_t integer_bits,
+                            int16_t* output) {
+  const int32_t int16_max = std::numeric_limits<int16_t>::max();
+  const int32_t int16_min = std::numeric_limits<int16_t>::min();
+  const double two = 2.0;
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      const float float_input =
+          input[index] * std::pow(two, static_cast<double>(integer_bits));
+      const float float_output = std::tanh(float_input);
+      const int32_t quant_output = static_cast<int32_t>(
+          float_output * static_cast<float>(std::pow(2, 15)));
+      const int32_t quant_output_clamped =
+          std::min(int16_max, std::max(int16_min, quant_output));
+      output[index] = static_cast<int16_t>(quant_output_clamped);
+    }
+  }
+}
+
+void PortableCwiseMul(const int16_t* input_1, const int16_t* input_2,
+                      int n_batch, int n_input, int shift, int16_t* output) {
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      const int16_t a = input_1[index];
+      const int16_t b = input_2[index];
+      const int32_t value = static_cast<int32_t>(a) * static_cast<int32_t>(b);
+      output[index] =
+          static_cast<int16_t>(gemmlowp::RoundingDivideByPOT(value, shift));
+    }
+  }
+}
+
+void PortableCwiseMul(const int16_t* input_1, const int16_t* input_2,
+                      int32_t multiplier, int32_t shift, int32_t n_batch,
+                      int32_t n_input, int32_t output_zp, int8_t* output) {
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      const int16_t a = input_1[index];
+      const int16_t b = input_2[index];
+      int32_t value = static_cast<int32_t>(a) * static_cast<int32_t>(b);
+      value = MultiplyByQuantizedMultiplier(value, multiplier, shift);
+      value -= output_zp;
+      value = std::min(std::max(static_cast<int32_t>(-128), value),
+                       static_cast<int32_t>(127));
+
+      output[index] = static_cast<int8_t>(value);
+    }
+  }
+}
+
+void PortableCwiseAdd(const int16_t* input_1, const int16_t* input_2,
+                      int n_batch, int n_input, int16_t* output) {
+  for (int batch = 0; batch < n_batch; ++batch) {
+    for (int i = 0; i < n_input; ++i) {
+      const int index = batch * n_input + i;
+      int32_t sum = input_1[index] + input_2[index];
+      const int32_t sum_clamped = std::min(kInt16Max, std::max(kInt16Min, sum));
+      output[index] = static_cast<int16_t>(sum_clamped);
+    }
+  }
+}
+
+float PortableVectorVectorDotProduct(const float* vector1, const float* vector2,
+                                     int v_size) {
+  float result = 0.0;
+  for (int v = 0; v < v_size; v++) {
+    result += *vector1++ * *vector2++;
+  }
+  return result;
+}
+
+namespace {
+inline int32_t VectorVectorDotProduct(const int16_t* vector1,
+                                      const int16_t* vector2, int v_size) {
+  int32_t result = 0;
+  for (int v = 0; v < v_size; v++) {
+    result += *vector1++ * *vector2++;
+  }
+  return result;
+}
+}  // namespace
+
+void PortableBatchVectorBatchVectorDotProduct(const int16_t* vector1,
+                                              const int16_t* vector2,
+                                              int v_size, int n_batch,
+                                              int32_t* result) {
+  for (int b = 0; b < n_batch; b++) {
+    result[b] = VectorVectorDotProduct(vector1, vector2, v_size);
+    vector1 += v_size;
+    vector2 += v_size;
+  }
+}
+
+void PortableVectorBatchVectorCwiseProductAccumulate(
+    const int16_t* vector, int v_size, const int16_t* batch_vector, int n_batch,
+    int32_t multiplier, int shift, int16_t* result) {
+  for (int b = 0; b < n_batch; b++) {
+    for (int v = 0; v < v_size; v++) {
+      int32_t prod = vector[v] * *batch_vector++;
+      prod = MultiplyByQuantizedMultiplier(prod, multiplier, shift);
+      int32_t output = prod + *result;
+      output = std::max(std::min(static_cast<int32_t>(32767), output),
+                        static_cast<int32_t>(-32768));
+      *result++ = output;
+    }
+  }
+}
+
+void PortableSub1Vector(const float* vector, int v_size, float* result) {
+  for (int v = 0; v < v_size; v++) {
+    *result++ = 1.0f - *vector++;
+  }
+}
+
+void PortableSub1Vector(const int16_t* vector, int v_size, int16_t* result) {
+  static const int16_t kOne = 32767;
+  for (int v = 0; v < v_size; v++) {
+    *result++ = kOne - *vector++;
+  }
+}
+
+void PortableVectorScalarMultiply(const int8_t* vector, const int v_size,
+                                  const float scale, float* result) {
+  for (int v = 0; v < v_size; ++v) {
+    *result++ = scale * *vector++;
+  }
+}
+
+void PortableMeanStddevNormalization(const float* __restrict__ input_vector,
+                                     float* __restrict__ output_vector,
+                                     int v_size, int n_batch) {
+  for (int batch = 0; batch < n_batch; ++batch) {
+    float sum = 0.0f;
+    for (int i = 0; i < v_size; ++i) {
+      sum += input_vector[i];
+    }
+    const float mean = sum / v_size;
+    float sum_diff_sq = 0.0f;
+    for (int i = 0; i < v_size; ++i) {
+      const float diff = input_vector[i] - mean;
+      sum_diff_sq += diff * diff;
+    }
+    const float variance = sum_diff_sq / v_size;
+    constexpr float kNormalizationConstant = 1e-8f;
+    const float stddev_inv =
+        1.0f / std::sqrt(variance + kNormalizationConstant);
+    for (int i = 0; i < v_size; ++i) {
+      output_vector[i] = (input_vector[i] - mean) * stddev_inv;
+    }
+    input_vector += v_size;
+    output_vector += v_size;
+  }
+}
+
+void PortableTwoGateSaturatingAdd(const int8_t* input, int8_t input_zp,
+                                  const int8_t* recurrent, int8_t recurrent_zp,
+                                  int32_t input_effective_scale_a,
+                                  int32_t input_effective_scale_b,
+                                  int32_t recurrent_effective_scale_a,
+                                  int32_t recurrent_effective_scale_b,
+                                  int32_t n_batch, int32_t n_cell,
+                                  int16_t* output) {
+  const int32_t int16_max = std::numeric_limits<int16_t>::max();
+  const int32_t int16_min = std::numeric_limits<int16_t>::min();
+  for (int i = 0; i < n_batch * n_cell; ++i) {
+    int32_t x = static_cast<int32_t>(input[i]) - static_cast<int32_t>(input_zp);
+    int32_t h =
+        static_cast<int32_t>(recurrent[i]) - static_cast<int32_t>(recurrent_zp);
+    int32_t x_scaled = MultiplyByQuantizedMultiplier(x, input_effective_scale_a,
+                                                     input_effective_scale_b);
+    int32_t h_scaled = MultiplyByQuantizedMultiplier(
+        h, recurrent_effective_scale_a, recurrent_effective_scale_b);
+    int32_t y = h_scaled + x_scaled;
+    if (y > int16_max) {
+      y = int16_max;
+    }
+    if (y < int16_min) {
+      y = int16_min;
+    }
+    output[i] = static_cast<int16_t>(y);
+  }
+}
+
+}  // namespace tensor_utils
+}  // namespace tflite

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h

@@ -0,0 +1,235 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_PORTABLE_TENSOR_UTILS_IMPL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PORTABLE_TENSOR_UTILS_IMPL_H_
+
+#include <algorithm>
+#include <cstdint>
+
+#if defined(_MSC_VER)
+#define __restrict__ __restrict
+#endif
+
+namespace tflite {
+
+// Not all backends support CpuBackendContext usage, so forward declare to avoid
+// pulling in its implementation.
+class CpuBackendContext;
+
+namespace tensor_utils {
+
+template <typename T>
+bool PortableIsZeroVector(const T* vector, int v_size) {
+  for (int i = 0; i < v_size; ++i) {
+    if (vector[i] != 0) {
+      return false;
+    }
+  }
+  return true;
+}
+
+void PortableSymmetricQuantizeFloats(const float* values, const int size,
+                                     int8_t* quantized_values, float* min_value,
+                                     float* max_value, float* scaling_factor);
+
+void PortableSymmetricQuantizeFloats(const float* values, const int size,
+                                     int8_t* quantized_values, float min_value,
+                                     float max_value, float* scaling_factor);
+
+void PortableAsymmetricQuantizeFloats(const float* values, const int size,
+                                      int8_t* quantized_values,
+                                      float* scaling_factor, int32_t* offset);
+
+// Multiply a matrix by a batch vector, and store results in a batch-size
+// vector.
+void PortableMatrixBatchVectorMultiplyAccumulate(const float* matrix,
+                                                 int m_rows, int m_cols,
+                                                 const float* vector,
+                                                 int n_batch, float* result);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vectors, const float* scaling_factors,
+    int n_batch, float* __restrict__ result);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vectors, const float* scaling_factors,
+    int n_batch, float* __restrict__ result, const float* per_channel_scale,
+    const int32_t* input_offset, int32_t* scratch, int32_t* row_sums,
+    bool* compute_row_sums, CpuBackendContext* context);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vector, const float* scaling_factors,
+    int n_batch, int32_t* scratch, float* __restrict__ result,
+    CpuBackendContext* context);
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate1x4(
+    const float* __restrict__ matrix, const int32_t* __restrict__ segments,
+    const int32_t* __restrict__ indices, int m_rows, int m_cols,
+    const float* __restrict__ vector, int n_batch, float* __restrict__ result);
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate(
+    const float* __restrict__ matrix, const uint8_t* __restrict__ ledger,
+    int m_rows, int m_cols, const float* __restrict__ vector, int n_batch,
+    float* __restrict__ result);
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const uint8_t* ledger, const int m_rows,
+    const int m_cols, const int8_t* __restrict__ vectors,
+    const float* scaling_factors, int n_batch, float* __restrict__ result);
+
+// Dot product of two vectors.
+float PortableVectorVectorDotProduct(const float* vector1, const float* vector2,
+                                     int v_size);
+
+void PortableBatchVectorBatchVectorDotProduct(const int16_t* vector1,
+                                              const int16_t* vector2,
+                                              int v_size, int n_batch,
+                                              int32_t* result);
+
+void PortableVectorBatchVectorCwiseProductAccumulate(
+    const int16_t* vector, int v_size, const int16_t* batch_vector, int n_batch,
+    int32_t multiplier, int shift, int16_t* result);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* input, const int32_t* bias,
+    const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
+    int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
+    int32_t* scratch, int16_t* output, CpuBackendContext* context);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* input, const int32_t* bias,
+    const int8_t* input_to_gate_weights, int32_t multiplier, int32_t shift,
+    int32_t n_batch, int32_t n_input, int32_t n_output, int32_t output_zp,
+    int32_t* scratch, int8_t* output, CpuBackendContext* context);
+
+void PortableMatrixBatchVectorMultiply(const int8_t* input,
+                                       int32_t input_zeropoint,
+                                       const int8_t* input_to_gate_weights,
+                                       int32_t input_to_gate_effective_scale_a,
+                                       int32_t input_to_gate_effective_scale_b,
+                                       int32_t n_batch, int32_t n_input,
+                                       int32_t n_cell, int8_t* gate_output,
+                                       int8_t gate_output_zp);
+
+void PortableMatrixBatchVectorMultiply(
+    const int16_t* hidden, const int8_t* hidden_to_output_weights,
+    int32_t proj_effective_scale_a, int32_t proj_effective_scale_b,
+    const int32_t* gate_bias, int32_t n_batch, int32_t n_hidden,
+    int32_t n_output, int32_t output_zp, int8_t* proj_output);
+
+void PortableMatrixScalarMultiplyAccumulate(const int8_t* matrix,
+                                            int32_t scalar, int32_t n_row,
+                                            int32_t n_col, int32_t* output);
+
+void PortableApplyLayerNorm(const int16_t* input,
+                            const int16_t* layer_norm_weights,
+                            const int32_t* bias, int32_t layer_norm_scale_a,
+                            int32_t layer_norm_scale_b, int32_t variance_limit,
+                            int n_batch, int n_input, int16_t* output);
+
+void PortableApplyLayerNormFloat(const int16_t* input,
+                                 const int16_t* layer_norm_weights,
+                                 int32_t layer_norm_scale_a,
+                                 int32_t layer_norm_scale_b,
+                                 const int32_t* bias, int n_batch, int n_input,
+                                 int16_t* output);
+
+void PortableApplySigmoid(const int16_t* input, int32_t n_batch,
+                          int32_t n_input, int16_t* output);
+
+void PortableApplySigmoidFloat(const int16_t* input, int32_t n_batch,
+                               int32_t n_input, int16_t* output);
+
+void PortableApplyTanh(int32_t integer_bits, const int16_t* input,
+                       int32_t n_batch, int32_t n_input, int16_t* output);
+
+void PortableApplyTanhFloat(const int16_t* input, int32_t n_batch,
+                            int32_t n_input, int32_t integer_bits,
+                            int16_t* output);
+
+void PortableCwiseMul(const int16_t* input_1, const int16_t* input_2,
+                      int n_batch, int n_input, int shift, int16_t* output);
+
+void PortableCwiseMul(const int16_t* input_1, const int16_t* input_2,
+                      int32_t multiplier, int32_t shift, int32_t n_batch,
+                      int32_t n_input, int32_t output_zp, int8_t* output);
+
+void PortableCwiseAdd(const int16_t* input_1, const int16_t* input_2,
+                      int n_batch, int n_input, int16_t* output);
+
+template <typename T>
+void PortableCwiseClipping(T* vector, const int v_size,
+                           const T& clipping_value) {
+  for (int i = 0; i < v_size; i++) {
+    vector[i] = std::max(std::min(clipping_value, vector[i]),
+                         static_cast<T>(-clipping_value));
+  }
+}
+
+// Batch vector initialization with another vector.
+void PortableVectorBatchVectorAssign(const float* vector, int v_size,
+                                     int n_batch, float* batch_vector);
+
+// Compute "1.0f - elements of vector" (used in CIFG).
+void PortableSub1Vector(const float* vector, int v_size, float* result);
+
+void PortableSub1Vector(const int16_t* vector, int v_size, int16_t* result);
+
+// Multiply all elements of vector with a scalar.
+void PortableVectorScalarMultiply(const int8_t* vector, int v_size, float scale,
+                                  float* result);
+
+// Reduce-sum on a vector:
+// input_vector: pointer to input vector.
+// output_vector: pointer to vector.
+// output_size: output vector size.
+// reduction_size: number of consecutive elements from input vector which are
+// added to get one element of output.
+template <typename INPUT, typename OUTPUT>
+void PortableReductionSumVector(const INPUT* input_vector,
+                                OUTPUT* output_vector, int output_size,
+                                int reduction_size) {
+  for (int o = 0; o < output_size; o++) {
+    OUTPUT result = 0;
+    for (int r = 0; r < reduction_size; r++) {
+      result += input_vector[r];
+    }
+    output_vector[o] = result;
+    input_vector += reduction_size;
+  }
+}
+
+// Layer norm for each batch.
+void PortableMeanStddevNormalization(const float* __restrict__ input_vector,
+                                     float* __restrict__ output_vector,
+                                     int v_size, int n_batch);
+
+// Saturate Add.
+void PortableTwoGateSaturatingAdd(const int8_t* input, int8_t input_zp,
+                                  const int8_t* recurrent, int8_t recurrent_zp,
+                                  int32_t input_effective_scale_a,
+                                  int32_t input_effective_scale_b,
+                                  int32_t recurrent_effective_scale_a,
+                                  int32_t recurrent_effective_scale_b,
+                                  int32_t n_batch, int32_t n_cell,
+                                  int16_t* output);
+
+}  // namespace tensor_utils
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PORTABLE_TENSOR_UTILS_IMPL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/prelu.h

@@ -0,0 +1,109 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_PRELU_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Broadcast prelu to output_shape for quantized uint8_t/int8_t data.
+template <typename T>
+inline void BroadcastPrelu4DSlow(
+    const PreluParams& params, const RuntimeShape& input_shape,
+    const T* input_data, const RuntimeShape& alpha_shape, const T* alpha_data,
+    const RuntimeShape& output_shape, T* output_data) {
+  TFLITE_DCHECK_LE(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(alpha_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), 4);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input_shape, alpha_shape, &desc1, &desc2);
+
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          int output_index = Offset(extended_output_shape, b, y, x, c);
+          int input_index = SubscriptToIndex(desc1, b, y, x, c);
+          const int32_t input_value =
+              params.input_offset + input_data[input_index];
+          int32_t output_value;
+          if (input_value >= 0) {
+            output_value = MultiplyByQuantizedMultiplier(
+                input_value, params.output_multiplier_1, params.output_shift_1);
+          } else {
+            auto alpha_index = SubscriptToIndex(desc2, b, y, x, c);
+            const int32_t alpha_value =
+                params.alpha_offset + alpha_data[alpha_index];
+
+            output_value = MultiplyByQuantizedMultiplier(
+                input_value * alpha_value, params.output_multiplier_2,
+                params.output_shift_2);
+          }
+          output_value += params.output_offset;
+
+          const int32_t quantized_min = std::numeric_limits<T>::min();
+          const int32_t quantized_max = std::numeric_limits<T>::max();
+          const int32_t clamped_output =
+              std::min(quantized_max, std::max(quantized_min, output_value));
+          output_data[output_index] = static_cast<T>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+inline void Prelu(const PreluParams& params, const RuntimeShape& input_shape,
+                  const T* input_data, const RuntimeShape& alpha_shape,
+                  const T* alpha_data, const RuntimeShape& output_shape,
+                  T* output_data) {
+  const int32_t quantized_min = std::numeric_limits<T>::min();
+  const int32_t quantized_max = std::numeric_limits<T>::max();
+
+  const int flat_size =
+      MatchingElementsSize(input_shape, alpha_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input_value = params.input_offset + input_data[i];
+    int32_t output_value;
+    if (input_value >= 0) {
+      output_value = MultiplyByQuantizedMultiplier(
+          input_value, params.output_multiplier_1, params.output_shift_1);
+    } else {
+      const int32_t alpha_value = params.alpha_offset + alpha_data[i];
+
+      output_value = MultiplyByQuantizedMultiplier(input_value * alpha_value,
+                                                   params.output_multiplier_2,
+                                                   params.output_shift_2);
+    }
+    output_value += params.output_offset;
+
+    const int32_t clamped_output =
+        std::min(quantized_max, std::max(quantized_min, output_value));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h

@@ -0,0 +1,138 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Consolidates dimensions in broadcast inputs, checks for five-fold pattern.
+//
+// For example, if sequence of dimensions of one input is
+// ..., 1, 3, 1, 7, 9, 5,... and the other is ..., 2, 3, 1, 7, 1, 1, ...
+// we can consolidate these as
+// ..., 1, 3*7, 9*5, ... and 2, 3*7, 1.
+//
+// The category is updated in the less-frequent case of shapes that are
+// not suited to a fivefold-loop broadcast.
+//
+// Falls back to generic pattern when it does not know how to process properly.
+//
+// Returns true iff there is some sort of broadcast, which includes five-fold
+// patterns and falling back to generic broadcast.
+inline bool ProcessBroadcastShapes(const RuntimeShape& shape0,
+                                   const RuntimeShape& shape1,
+                                   tflite::ArithmeticParams* params) {
+  const int dims_count =
+      std::max(shape0.DimensionsCount(), shape1.DimensionsCount());
+
+  params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast;
+  RuntimeShape scalar_shape(dims_count, 1);
+
+  auto extended_shape0 = RuntimeShape::ExtendedShape(dims_count, shape0);
+  auto extended_shape1 = RuntimeShape::ExtendedShape(dims_count, shape1);
+
+  // Check for "exact" match, implicitly accepting any scalar shapes.
+  if (extended_shape0 == extended_shape1) {
+    params->broadcast_category = BroadcastableOpCategory::kNonBroadcast;
+    return false;
+  }
+
+  for (int i = dims_count - 1; i >= 0; --i) {
+    if (extended_shape0.Dims(i) == extended_shape1.Dims(i)) {
+      continue;
+    } else if (extended_shape0.Dims(i) == 1) {
+      params->broadcast_category =
+          BroadcastableOpCategory::kFirstInputBroadcastsFast;
+      break;
+    } else if (extended_shape1.Dims(i) == 1) {
+      params->broadcast_category =
+          BroadcastableOpCategory::kSecondInputBroadcastsFast;
+      break;
+    } else {
+      // This case is erroneous: there is a dimension that does not match and
+      // is not a broadcast from one shape to the other.
+      params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast;
+      return true;
+    }
+  }
+
+  if (params->broadcast_category !=
+          BroadcastableOpCategory::kFirstInputBroadcastsFast &&
+      params->broadcast_category !=
+          BroadcastableOpCategory::kSecondInputBroadcastsFast) {
+    // This is unreachable because at least one else clause in the above loop
+    // must be reached.
+    TFLITE_DCHECK(false);
+    params->broadcast_category = BroadcastableOpCategory::kNonBroadcast;
+    return false;
+  }
+
+  // From this point it is assumed contractually that corresponding dimensions
+  // in shape0 and shape1 are either (a) equal or (b) one or other equals 1.
+  const bool swap_inputs = params->broadcast_category ==
+                           BroadcastableOpCategory::kSecondInputBroadcastsFast;
+  const RuntimeShape* shape_a =
+      swap_inputs ? &extended_shape1 : &extended_shape0;
+  const RuntimeShape* shape_b =
+      swap_inputs ? &extended_shape0 : &extended_shape1;
+
+  int i = dims_count - 1;
+  params->broadcast_shape[0] = 1;
+  params->broadcast_shape[1] = 1;
+  params->broadcast_shape[2] = 1;
+  params->broadcast_shape[3] = 1;
+  params->broadcast_shape[4] = 1;
+  // y_0 is greedy: include dims if both or neither equal 1: in other words,
+  // test for equality rather than (shape_a->Dims(i) != 1).
+  while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) {
+    params->broadcast_shape[4] *= shape_b->Dims(i);
+    --i;
+  }
+  // Here either input_a or input_b has dim of 1 (if i >= 0).  If it is input_b
+  // that has the unit dimension, the next two loops are not entered.
+  while (i >= 0 && shape_a->Dims(i) == 1) {
+    params->broadcast_shape[3] *= shape_b->Dims(i);
+    --i;
+  }
+  while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) {
+    params->broadcast_shape[2] *= shape_a->Dims(i);
+    --i;
+  }
+  // Here either input_a or input_b has dim of 1 (if i >= 0).
+  while (i >= 0 && shape_b->Dims(i) == 1) {
+    params->broadcast_shape[1] *= shape_a->Dims(i);
+    --i;
+  }
+  while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) {
+    params->broadcast_shape[0] *= shape_b->Dims(i);
+    --i;
+  }
+
+  // Rarer case is when the broadcast dimensions cannot be handled by a fivefold
+  // loop.
+  if (i >= 0) {
+    params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast;
+  }
+  return true;
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/quantize.h

@@ -0,0 +1,89 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_
+
+#include <algorithm>
+#include <limits>
+#include <vector>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename InputT, typename OutputT>
+inline void AffineQuantize(const tflite::QuantizationParams& op_params,
+                           const RuntimeShape& input_shape,
+                           const InputT* input_data,
+                           const RuntimeShape& output_shape,
+                           OutputT* output_data) {
+  const int32_t zero_point = op_params.zero_point;
+  const double scale = op_params.scale;
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  static constexpr int32_t min_val = std::numeric_limits<OutputT>::min();
+  static constexpr int32_t max_val = std::numeric_limits<OutputT>::max();
+
+  for (int i = 0; i < flat_size; i++) {
+    const InputT val = input_data[i];
+    int32_t unclamped =
+        static_cast<int32_t>(TfLiteRound(val / static_cast<float>(scale))) +
+        zero_point;
+    int32_t clamped = std::min(std::max(unclamped, min_val), max_val);
+    output_data[i] = clamped;
+  }
+}
+
+// Quantizes per-channel.
+template <typename InputT, typename OutputT>
+inline void PerChannelQuantize(
+    const tflite::PerChannelQuantizationParams& op_params,
+    const RuntimeShape& input_shape, const InputT* input_data,
+    const RuntimeShape& output_shape, OutputT* output_data) {
+  // Ensure flat size is same.
+  MatchingFlatSize(input_shape, output_shape);
+
+  const int32_t* zero_point = op_params.zero_point;
+  const float* scale = op_params.scale;
+  const int32_t quantized_dimension = op_params.quantized_dimension;
+  const int32_t num_dims = input_shape.DimensionsCount();
+  const int32_t* dims_data = input_shape.DimsData();
+  std::vector<int> current_dim(num_dims, 0);
+  static constexpr int32_t min_val = std::numeric_limits<OutputT>::min();
+  static constexpr int32_t max_val = std::numeric_limits<OutputT>::max();
+
+  do {
+    size_t offset =
+        ReducedOutputOffset(num_dims, reinterpret_cast<const int*>(dims_data),
+                            current_dim.data(), 0, nullptr);
+    const InputT val = input_data[offset];
+    const int channel = current_dim[quantized_dimension];
+    int32_t unclamped = static_cast<int32_t>(TfLiteRound(
+                            val / static_cast<float>(scale[channel]))) +
+                        zero_point[channel];
+    int32_t clamped = std::min(std::max(unclamped, min_val), max_val);
+    output_data[offset] = static_cast<OutputT>(clamped);
+  } while (NextIndex(num_dims, reinterpret_cast<const int*>(dims_data),
+                     current_dim.data()));
+}
+
+}  // namespace reference_ops
+
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/reduce.h

@@ -0,0 +1,524 @@
+/* Copyright 2019 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_KERNELS_INTERNAL_REFERENCE_REDUCE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REDUCE_H_
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/max.h"
+#include "tensorflow/lite/kernels/internal/min.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+// Check if the reduction at index is the first one along the dimensions given
+// in axis.
+inline bool IsFirstReduction(const int* index, const int num_axis,
+                             const int* axis) {
+  if (num_axis == 0) {
+    return true;
+  }
+
+  TFLITE_DCHECK(index != nullptr);
+  TFLITE_DCHECK(axis != nullptr);
+  for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) {
+    if (index[axis[axis_idx]] != 0) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+namespace tflite {
+
+namespace reference_ops {
+
+// A generic reduce method that can be used for reduce_sum, reduce_mean, etc.
+// This method iterates through input data and reduce elements along the
+// dimensions given in axis.
+template <typename In, typename Out>
+inline bool Reduce(const In* input_data, const int* input_dims,
+                   const int* output_dims, const int input_num_dims,
+                   const int output_num_dims, const int* axis,
+                   const int num_axis, int* input_iter,
+                   Out reducer(Out current, const In in), Out* output_data) {
+  // Reset input iterator.
+  for (int idx = 0; idx < input_num_dims; ++idx) {
+    input_iter[idx] = 0;
+  }
+  // Iterate through input_data.
+  do {
+    size_t input_offset =
+        ReducedOutputOffset(input_num_dims, input_dims, input_iter, 0, nullptr);
+    size_t output_offset = ReducedOutputOffset(input_num_dims, input_dims,
+                                               input_iter, num_axis, axis);
+    output_data[output_offset] =
+        reducer(output_data[output_offset], input_data[input_offset]);
+  } while (NextIndex(input_num_dims, input_dims, input_iter));
+  return true;
+}
+
+// Similar to above Reduce function but takes two reducer functions.
+// The 'reducer_first' is called with the first value of the reduction,
+// 'reducer_next' is then called for all the others.
+template <typename In, typename Out>
+inline bool Reduce(const In* input_data, const int* input_dims,
+                   const int* output_dims, const int input_num_dims,
+                   const int output_num_dims, const int* axis,
+                   const int num_axis, int* input_iter,
+                   const std::function<Out(In in)>& reducer_first,
+                   const std::function<Out(Out current, In in)>& reducer_next,
+                   Out* output_data) {
+  // Reset input iterator.
+  for (int idx = 0; idx < input_num_dims; ++idx) {
+    input_iter[idx] = 0;
+  }
+  // Iterate through input_data.
+  do {
+    size_t input_offset =
+        ReducedOutputOffset(input_num_dims, input_dims, input_iter, 0, nullptr);
+    size_t output_offset = ReducedOutputOffset(input_num_dims, input_dims,
+                                               input_iter, num_axis, axis);
+    if (IsFirstReduction(input_iter, num_axis, axis)) {
+      output_data[output_offset] = reducer_first(input_data[input_offset]);
+    } else {
+      output_data[output_offset] =
+          reducer_next(output_data[output_offset], input_data[input_offset]);
+    }
+  } while (NextIndex(input_num_dims, input_dims, input_iter));
+  return true;
+}
+
+// This method parses the input 'axis' to remove duplicates and handle negative
+// values, and returns a valid 'out_axis'
+inline bool ResolveAxis(const int num_dims, const int* axis,
+                        const int64_t num_axis, int* out_axis,
+                        int* out_num_axis) {
+  *out_num_axis = 0;  // Just in case.
+  // Short-circuit axis resolution for scalars; the axis will go unused.
+  if (num_dims == 0) {
+    return true;
+  }
+  // o(n^2) is fine since out_num_axis should be really small, mostly <= 4
+  for (int64_t idx = 0; idx < num_axis; ++idx) {
+    // Handle negative index. A positive index 'p_idx' can be represented as a
+    // negative index 'n_idx' as: n_idx = p_idx-num_dims
+    // eg: For num_dims=3, [0, 1, 2] is the same as [-3, -2, -1]  */
+    int current = axis[idx] < 0 ? (axis[idx] + num_dims) : axis[idx];
+    TFLITE_DCHECK(current >= 0 && current < num_dims);
+    if (current < 0 || current >= num_dims) {
+      return false;
+    }
+    bool is_dup = false;
+    for (int j = 0; j < *out_num_axis; ++j) {
+      if (out_axis[j] == current) {
+        is_dup = true;
+        break;
+      }
+    }
+    if (!is_dup) {
+      out_axis[*out_num_axis] = current;
+      *out_num_axis += 1;
+    }
+  }
+  return true;
+}
+
+// This method expects that output_data has been initialized.
+template <typename In, typename Out>
+inline bool ReduceSumImpl(const In* input_data, const int* input_dims,
+                          const int* output_dims, const int input_num_dims,
+                          const int output_num_dims, const int* axis,
+                          const int num_axis, int* input_iter,
+                          Out* output_data) {
+  auto reducer = [](const Out current, const In in) -> Out {
+    const Out actual_in = static_cast<Out>(in);
+    return current + actual_in;
+  };
+  return Reduce<In, Out>(input_data, input_dims, output_dims, input_num_dims,
+                         output_num_dims, axis, num_axis, input_iter, reducer,
+                         output_data);
+}
+
+template <typename T>
+inline bool InitTensorDataForReduce(const int* dims, const int num_dims,
+                                    const T init_value, T* data) {
+  size_t num_elements = 1;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    size_t current = static_cast<size_t>(dims[idx]);
+    // Overflow prevention.
+    if (current > 0 &&
+        num_elements > std::numeric_limits<size_t>::max() / current) {
+      return false;
+    }
+    num_elements *= current;
+  }
+  for (size_t idx = 0; idx < num_elements; ++idx) {
+    data[idx] = init_value;
+  }
+  return true;
+}
+
+// Computes the generic value (i.e., sum/max/min/prod) of elements across
+// dimensions given in axis. It needs to pass in init_value and reducer.
+template <typename T>
+inline bool ReduceGeneric(const T* input_data, const int* input_dims,
+                          const int input_num_dims, T* output_data,
+                          const int* output_dims, const int output_num_dims,
+                          const int* axis, const int64_t num_axis_dimensions,
+                          bool keep_dims, int* temp_index, int* resolved_axis,
+                          T init_value,
+                          T reducer(const T current, const T in)) {
+  // Reset output data.
+  if (!InitTensorDataForReduce(output_dims, output_num_dims, init_value,
+                               output_data)) {
+    return false;
+  }
+
+  // Return early when input shape has zero dim. This is done after initializing
+  // data for output tensor because there are cases that the input tensor is
+  // empty but output tensor is not. In that case, output tensor should be
+  // filled with init_value.
+  for (int i = 0; i < input_num_dims; ++i) {
+    if (input_dims[i] == 0) return true;
+  }
+
+  // Resolve axis.
+  int num_resolved_axis = 0;
+  if (!ResolveAxis(input_num_dims, axis, num_axis_dimensions, resolved_axis,
+                   &num_resolved_axis)) {
+    return false;
+  }
+
+  return Reduce<T, T>(input_data, input_dims, output_dims, input_num_dims,
+                      output_num_dims, resolved_axis, num_resolved_axis,
+                      temp_index, reducer, output_data);
+}
+
+// Computes the mean of elements across dimensions given in axis.
+// It does so in two stages, first calculates the sum of elements along the axis
+// then divides it by the number of element in axis.
+template <typename T, typename U>
+inline bool Mean(const T* input_data, const int* input_dims,
+                 const int input_num_dims, T* output_data,
+                 const int* output_dims, const int output_num_dims,
+                 const int* axis, const int num_axis_dimensions, bool keep_dims,
+                 int* temp_index, int* resolved_axis, U* temp_sum) {
+  ruy::profiler::ScopeLabel label("Mean");
+  // Reset output data.
+  size_t num_outputs = 1;
+  for (int idx = 0; idx < output_num_dims; ++idx) {
+    size_t current = static_cast<size_t>(output_dims[idx]);
+    // Overflow prevention.
+    if (num_outputs > std::numeric_limits<size_t>::max() / current) {
+      return false;
+    }
+    num_outputs *= current;
+  }
+  for (size_t idx = 0; idx < num_outputs; ++idx) {
+    output_data[idx] = T();
+    temp_sum[idx] = U();
+  }
+
+  // Resolve axis.
+  int num_resolved_axis = 0;
+  if (!ResolveAxis(input_num_dims, axis, num_axis_dimensions, resolved_axis,
+                   &num_resolved_axis)) {
+    return false;
+  }
+
+  if (!ReduceSumImpl<T, U>(input_data, input_dims, output_dims, input_num_dims,
+                           output_num_dims, resolved_axis, num_resolved_axis,
+                           temp_index, temp_sum)) {
+    return false;
+  }
+
+  // Calculate mean by dividing output_data by num of aggregated element.
+  size_t num_elements_in_axis = 1;
+  for (int idx = 0; idx < num_resolved_axis; ++idx) {
+    size_t current = static_cast<size_t>(input_dims[resolved_axis[idx]]);
+    // Overflow prevention.
+    if (current > (std::numeric_limits<size_t>::max() / num_elements_in_axis)) {
+      return false;
+    }
+    num_elements_in_axis *= current;
+  }
+
+  if (num_elements_in_axis > 0) {
+    for (size_t idx = 0; idx < num_outputs; ++idx) {
+      output_data[idx] =
+          static_cast<T>(temp_sum[idx] / static_cast<U>(num_elements_in_axis));
+    }
+  }
+  return true;
+}
+
+template <typename T>
+inline void Mean(const tflite::MeanParams& op_params,
+                 const RuntimeShape& unextended_input_shape,
+                 const T* input_data,
+                 const RuntimeShape& unextended_output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("Mean4D");
+
+  // Current implementation only supports dimension equals 4 and simultaneous
+  // reduction over width and height.
+  TFLITE_CHECK_EQ(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_CHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  const int output_batch = output_shape.Dims(0);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int output_depth = output_shape.Dims(3);
+
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+
+  TFLITE_CHECK_EQ(op_params.axis_count, 2);
+  TFLITE_CHECK((op_params.axis[0] == 1 && op_params.axis[1] == 2) ||
+               (op_params.axis[0] == 2 && op_params.axis[1] == 1));
+  TFLITE_CHECK_EQ(output_height, 1);
+  TFLITE_CHECK_EQ(output_width, 1);
+
+  for (int out_b = 0; out_b < output_batch; ++out_b) {
+    for (int out_d = 0; out_d < output_depth; ++out_d) {
+      float value = 0;
+      for (int in_h = 0; in_h < input_height; ++in_h) {
+        for (int in_w = 0; in_w < input_width; ++in_w) {
+          value += input_data[Offset(input_shape, out_b, in_h, in_w, out_d)];
+        }
+      }
+      output_data[Offset(output_shape, out_b, 0, 0, out_d)] =
+          value / (input_width * input_height);
+    }
+  }
+}
+
+inline void Mean(const tflite::MeanParams& op_params,
+                 const RuntimeShape& unextended_input_shape,
+                 const uint8_t* input_data, int32_t input_zero_point,
+                 float input_scale, const RuntimeShape& unextended_output_shape,
+                 uint8_t* output_data, int32_t output_zero_point,
+                 float output_scale) {
+  ruy::profiler::ScopeLabel label("Mean4D/Uint8");
+
+  // Current implementation only supports dimension equals 4 and simultaneous
+  // reduction over width and height.
+  TFLITE_CHECK_EQ(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_CHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+  const int output_batch = output_shape.Dims(0);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int output_depth = output_shape.Dims(3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const float num_elements_in_axis = input_width * input_height;
+
+  TFLITE_CHECK_EQ(op_params.axis_count, 2);
+  TFLITE_CHECK((op_params.axis[0] == 1 && op_params.axis[1] == 2) ||
+               (op_params.axis[0] == 2 && op_params.axis[1] == 1));
+  TFLITE_CHECK_EQ(output_height, 1);
+  TFLITE_CHECK_EQ(output_width, 1);
+
+  constexpr int32_t kMinValue = std::numeric_limits<uint8_t>::min();
+  constexpr int32_t kMaxValue = std::numeric_limits<uint8_t>::max();
+
+  float temp = input_zero_point * input_scale / output_scale;
+  temp = temp > 0 ? temp + 0.5f : temp - 0.5f;
+  int32_t bias = output_zero_point - static_cast<int32_t>(temp);
+  double real_scale =
+      static_cast<double>(input_scale / (num_elements_in_axis * output_scale));
+
+  int32_t multiplier;
+  int shift;
+  QuantizeMultiplier(real_scale, &multiplier, &shift);
+  for (int out_b = 0; out_b < output_batch; ++out_b) {
+    for (int out_d = 0; out_d < output_depth; ++out_d) {
+      int32_t acc = 0;
+      for (int in_h = 0; in_h < input_height; ++in_h) {
+        for (int in_w = 0; in_w < input_width; ++in_w) {
+          acc += input_data[Offset(input_shape, out_b, in_h, in_w, out_d)];
+        }
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, multiplier, shift);
+      acc += bias;
+      acc = std::min(std::max(acc, kMinValue), kMaxValue);
+      output_data[Offset(output_shape, out_b, 0, 0, out_d)] =
+          static_cast<uint8_t>(acc);
+    }
+  }
+}
+
+// Computes the mean of elements across dimensions given in axis.
+// It does so in two stages, first calculates the sum of elements along the axis
+// then divides it by the number of element in axis for quantized values.
+template <typename T, typename U>
+inline bool QuantizedMeanOrSum(const T* input_data, int32_t input_zero_point,
+                               float input_scale, const int* input_dims,
+                               const int input_num_dims, T* output_data,
+                               int32_t output_zero_point, float output_scale,
+                               const int* output_dims,
+                               const int output_num_dims, const int* axis,
+                               const int num_axis_dimensions, bool keep_dims,
+                               int* temp_index, int* resolved_axis, U* temp_sum,
+                               bool compute_sum) {
+  const bool uint8_case = std::is_same<T, uint8_t>::value;
+  const bool int16_case = std::is_same<T, int16_t>::value;
+  if (uint8_case) {
+    ruy::profiler::ScopeLabel label(compute_sum ? "Sum/Uint8" : "Mean/Uint8");
+  } else if (int16_case) {
+    ruy::profiler::ScopeLabel label(compute_sum ? "Sum/Int16" : "Mean/Int16");
+  } else {
+    ruy::profiler::ScopeLabel label(compute_sum ? "Sum/Int8" : "Mean/Int8");
+  }
+  // Reset output data.
+  size_t num_outputs = 1;
+  for (int idx = 0; idx < output_num_dims; ++idx) {
+    size_t current = static_cast<size_t>(output_dims[idx]);
+    // Overflow prevention.
+    if (num_outputs > std::numeric_limits<size_t>::max() / current) {
+      return false;
+    }
+    num_outputs *= current;
+  }
+  for (size_t idx = 0; idx < num_outputs; ++idx) {
+    output_data[idx] = T();
+    temp_sum[idx] = U();
+  }
+
+  // Return early when input shape has zero dim. This is done after initializing
+  // data for output tensor because there are cases that the input tensor is
+  // empty but output tensor is not. In that case, output tensor should be
+  // filled with init_value.
+  for (int i = 0; i < input_num_dims; ++i) {
+    if (input_dims[i] == 0) return true;
+  }
+
+  // Resolve axis.
+  int num_resolved_axis = 0;
+  if (!ResolveAxis(input_num_dims, axis, num_axis_dimensions, resolved_axis,
+                   &num_resolved_axis)) {
+    return false;
+  }
+
+  if (!ReduceSumImpl<T, U>(input_data, input_dims, output_dims, input_num_dims,
+                           output_num_dims, resolved_axis, num_resolved_axis,
+                           temp_index, temp_sum)) {
+    return false;
+  }
+
+  // Calculate mean by dividing output_data by num of aggregated element.
+  size_t num_elements_in_axis = 1;
+  for (int idx = 0; idx < num_resolved_axis; ++idx) {
+    size_t current = static_cast<size_t>(input_dims[resolved_axis[idx]]);
+    // Overflow prevention.
+    if (current > (std::numeric_limits<size_t>::max() / num_elements_in_axis)) {
+      return false;
+    }
+    num_elements_in_axis *= current;
+  }
+
+  if (num_elements_in_axis > 0) {
+    const float scale = input_scale / output_scale;
+    if (compute_sum) {
+      // TODO(b/116341117): Eliminate float and do this completely in 8bit.
+      const float bias = -input_zero_point * scale * num_elements_in_axis;
+      for (size_t idx = 0; idx < num_outputs; ++idx) {
+        const U value =
+            static_cast<U>(TfLiteRound(temp_sum[idx] * scale + bias)) +
+            output_zero_point;
+        output_data[idx] = static_cast<T>(value);
+      }
+    } else {
+      const float bias = -input_zero_point * scale;
+      for (size_t idx = 0; idx < num_outputs; ++idx) {
+        float float_mean = static_cast<float>(temp_sum[idx]) /
+                           static_cast<float>(num_elements_in_axis);
+        float result = TfLiteMin(
+            TfLiteRound(float_mean * scale + bias) + output_zero_point,
+            static_cast<float>(std::numeric_limits<T>::max()));
+        result = TfLiteMax(result,
+                           static_cast<float>(std::numeric_limits<T>::min()));
+        output_data[idx] = static_cast<T>(result);
+      }
+    }
+  }
+  return true;
+}
+
+template <typename T>
+inline bool QuantizedReduceProd(const T* input_data, int32_t input_zero_point,
+                                const RuntimeShape& input_shape, T* output_data,
+                                int32_t output_zero_point,
+                                const RuntimeShape& output_shape,
+                                const int* axis,
+                                const int64_t num_axis_dimensions,
+                                bool keep_dims, int* temp_index,
+                                int* resolved_axis, int32_t* temp_prod,
+                                int32_t scaling_multiplier, int scaling_shift) {
+  const int32_t kMinValue = std::numeric_limits<T>::min();
+  const int32_t kMaxValue = std::numeric_limits<T>::max();
+
+  // Resolve axis.
+  int num_resolved_axis = 0;
+  if (!ResolveAxis(input_shape.DimensionsCount(), axis, num_axis_dimensions,
+                   resolved_axis, &num_resolved_axis)) {
+    return false;
+  }
+
+  // Calculate the reduced product by rescaling each multiplication step to
+  // avoid an overflow.
+  auto reducer_first = [&](T in) -> int32_t { return in - input_zero_point; };
+
+  auto reducer_next = [&](int32_t current, T in) -> int32_t {
+    const int64_t result =
+        static_cast<int64_t>(current) * (in - input_zero_point);
+    return MultiplyByQuantizedMultiplier(result, scaling_multiplier,
+                                         scaling_shift);
+  };
+
+  if (!Reduce<T, int32_t>(
+          input_data, input_shape.DimsData(), output_shape.DimsData(),
+          input_shape.DimensionsCount(), output_shape.DimensionsCount(),
+          resolved_axis, num_resolved_axis, temp_index, reducer_first,
+          reducer_next, temp_prod)) {
+    return false;
+  }
+
+  for (int i = 0; i < output_shape.FlatSize(); i++) {
+    int32_t result =
+        MultiplyByQuantizedMultiplier(static_cast<int64_t>(temp_prod[i]),
+                                      scaling_multiplier, scaling_shift) +
+        output_zero_point;
+    result = std::min(std::max(result, kMinValue), kMaxValue);
+    output_data[i] = static_cast<T>(result);
+  }
+
+  return true;
+}
+
+}  // namespace reference_ops
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REDUCE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/requantize.h

@@ -0,0 +1,68 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename input_type, typename output_type>
+inline void Requantize(const input_type* input_data, int32_t size,
+                       int32_t effective_scale_multiplier,
+                       int32_t effective_scale_shift, int32_t input_zeropoint,
+                       int32_t output_zeropoint, output_type* output_data) {
+  ruy::profiler::ScopeLabel label("Requantize");
+  const bool same_scale =
+      (effective_scale_multiplier == 1 << 30 && effective_scale_shift == 1);
+  if (same_scale) {
+    const bool mixed_type_int8_uint8 =
+        std::is_same<input_type, int8_t>::value &&
+        std::is_same<output_type, uint8_t>::value;
+    const bool mixed_type_uint8_int8 =
+        std::is_same<input_type, uint8_t>::value &&
+        std::is_same<output_type, int8_t>::value;
+    const int32_t zero_point_diff = input_zeropoint - output_zeropoint;
+    // Fast path to do requantization for the case when just a shift of 128 is
+    // needed.
+    if ((mixed_type_int8_uint8 && zero_point_diff == -128) ||
+        (mixed_type_uint8_int8 && zero_point_diff == 128)) {
+      for (int i = 0; i < size; ++i) {
+        output_data[i] = input_data[i] ^ 0x80;
+      }
+      return;
+    }
+  }
+  static constexpr int32_t kMinOutput = std::numeric_limits<output_type>::min();
+  static constexpr int32_t kMaxOutput = std::numeric_limits<output_type>::max();
+  for (int i = 0; i < size; ++i) {
+    const int32_t input = input_data[i] - input_zeropoint;
+    const int32_t output =
+        MultiplyByQuantizedMultiplier(input, effective_scale_multiplier,
+                                      effective_scale_shift) +
+        output_zeropoint;
+    const int32_t clamped_output =
+        std::max(std::min(output, kMaxOutput), kMinOutput);
+    output_data[i] = static_cast<output_type>(clamped_output);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/resize_bilinear.h

@@ -0,0 +1,228 @@
+/* Copyright 2021 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_KERNELS_INTERNAL_REFERENCE_RESIZE_BILINEAR_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_BILINEAR_H_
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void ComputeInterpolationValues(const float value, const float scale,
+                                       const bool half_pixel_centers,
+                                       int32_t input_size, float* scaled_value,
+                                       int32_t* lower_bound,
+                                       int32_t* upper_bound) {
+  if (half_pixel_centers) {
+    *scaled_value = (value + 0.5f) * scale - 0.5f;
+  } else {
+    *scaled_value = value * scale;
+  }
+  float scaled_value_floor = std::floor(*scaled_value);
+  *lower_bound = std::max(static_cast<int32_t>(scaled_value_floor),
+                          static_cast<int32_t>(0));
+  *upper_bound =
+      std::min(static_cast<int32_t>(std::ceil(*scaled_value)), input_size - 1);
+}
+
+template <typename T>
+inline void ResizeBilinear(const tflite::ResizeBilinearParams& op_params,
+                           const RuntimeShape& unextended_input_shape,
+                           const T* input_data,
+                           const RuntimeShape& unextended_output_size_shape,
+                           const int32_t* output_size_data,
+                           const RuntimeShape& unextended_output_shape,
+                           T* output_data) {
+  // If half_pixel_centers is True, align_corners must be False.
+  TFLITE_DCHECK(!op_params.half_pixel_centers || !op_params.align_corners);
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_size_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_size_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_size_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  int32_t batches = MatchingDim(input_shape, 0, output_shape, 0);
+  int32_t input_height = input_shape.Dims(1);
+  int32_t input_width = input_shape.Dims(2);
+  int32_t depth = MatchingDim(input_shape, 3, output_shape, 3);
+
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(0), 1);
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(1), 1);
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(2), 1);
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(3), 2);
+  int32_t output_height =
+      output_size_data[Offset(output_size_shape, 0, 0, 0, 0)];
+  int32_t output_width =
+      output_size_data[Offset(output_size_shape, 0, 0, 0, 1)];
+
+  float height_scale = static_cast<float>(input_height) / output_height;
+  float width_scale = static_cast<float>(input_width) / output_width;
+  if (op_params.align_corners && output_height > 1) {
+    height_scale = static_cast<float>(input_height - 1) / (output_height - 1);
+  }
+  if (op_params.align_corners && output_width > 1) {
+    width_scale = static_cast<float>(input_width - 1) / (output_width - 1);
+  }
+  const float rounding_offset = std::numeric_limits<T>::is_integer ? .5f : .0f;
+
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < output_height; ++y) {
+      float input_y;
+      int32_t y0, y1;
+      ComputeInterpolationValues(y, height_scale, op_params.half_pixel_centers,
+                                 input_height, &input_y, &y0, &y1);
+      for (int x = 0; x < output_width; ++x) {
+        float input_x;
+        int32_t x0, x1;
+        ComputeInterpolationValues(x, width_scale, op_params.half_pixel_centers,
+                                   input_width, &input_x, &x0, &x1);
+        for (int c = 0; c < depth; ++c) {
+          T interpolation =
+              static_cast<T>(input_data[Offset(input_shape, b, y0, x0, c)] *
+                                 (1 - (input_y - y0)) * (1 - (input_x - x0)) +
+                             input_data[Offset(input_shape, b, y1, x0, c)] *
+                                 (input_y - y0) * (1 - (input_x - x0)) +
+                             input_data[Offset(input_shape, b, y0, x1, c)] *
+                                 (1 - (input_y - y0)) * (input_x - x0) +
+                             input_data[Offset(input_shape, b, y1, x1, c)] *
+                                 (input_y - y0) * (input_x - x0) +
+                             rounding_offset);
+          output_data[Offset(output_shape, b, y, x, c)] = interpolation;
+        }
+      }
+    }
+  }
+}
+
+inline void ComputeInterpolationValuesInteger(
+    const int32_t value, const int32_t scale_10, const bool half_pixel_centers,
+    int32_t input_size, int32_t* scaled_value, int32_t* lower_bound,
+    int32_t* upper_bound) {
+  if (half_pixel_centers) {
+    *scaled_value = value * scale_10 + scale_10 / 2 - (1 << 9);
+  } else {
+    *scaled_value = value * scale_10;
+  }
+  constexpr int32_t zero = 0;
+  *lower_bound = std::max(*scaled_value / (1 << 10), zero);
+  *upper_bound =
+      std::min((*scaled_value + (1 << 10) - 1) / (1 << 10), input_size - 1);
+}
+
+// Same as above but doesn't use any floating-point for the resize
+template <typename T>
+inline void ResizeBilinearInteger(
+    const tflite::ResizeBilinearParams& op_params,
+    const RuntimeShape& unextended_input_shape, const T* input_data,
+    const RuntimeShape& unextended_output_size_shape,
+    const int32_t* output_size_data,
+    const RuntimeShape& unextended_output_shape, T* output_data) {
+  // If half_pixel_centers is True, align_corners must be False.
+  TFLITE_DCHECK(!op_params.half_pixel_centers || !op_params.align_corners);
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_size_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_size_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_size_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  const int32_t batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int32_t input_height = input_shape.Dims(1);
+  const int32_t input_width = input_shape.Dims(2);
+  const int32_t depth = MatchingDim(input_shape, 3, output_shape, 3);
+
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(0), 1);
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(1), 1);
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(2), 1);
+  TFLITE_DCHECK_EQ(output_size_shape.Dims(3), 2);
+  const int32_t output_height =
+      output_size_data[Offset(output_size_shape, 0, 0, 0, 0)];
+  const int32_t output_width =
+      output_size_data[Offset(output_size_shape, 0, 0, 0, 1)];
+
+  int32_t height_scale_10 =
+      ((1 << 10) * input_height + output_height / 2) / output_height;
+  int32_t width_scale_10 =
+      ((1 << 10) * input_width + output_width / 2) / output_width;
+  if (op_params.align_corners && output_height > 1) {
+    height_scale_10 =
+        ((1 << 10) * (input_height - 1) + (output_height - 1) / 2) /
+        (output_height - 1);
+  }
+  if (op_params.align_corners && output_width > 1) {
+    width_scale_10 = ((1 << 10) * (input_width - 1) + (output_width - 1) / 2) /
+                     (output_width - 1);
+  }
+
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < output_height; ++y) {
+      int32_t input_y, y0, y1;
+      ComputeInterpolationValuesInteger(y, height_scale_10,
+                                        op_params.half_pixel_centers,
+                                        input_height, &input_y, &y0, &y1);
+      for (int x = 0; x < output_width; ++x) {
+        int32_t input_x, x0, x1;
+        ComputeInterpolationValuesInteger(x, width_scale_10,
+                                          op_params.half_pixel_centers,
+                                          input_width, &input_x, &x0, &x1);
+        for (int c = 0; c < depth; ++c) {
+          const int64_t output_20_ll =
+              static_cast<int64_t>(
+                  input_data[Offset(input_shape, b, y0, x0, c)]) *
+              ((1 << 10) - (input_y - (1 << 10) * y0)) *
+              ((1 << 10) - (input_x - (1 << 10) * x0));
+          const int64_t output_20_lu =
+              static_cast<int64_t>(
+                  input_data[Offset(input_shape, b, y1, x0, c)]) *
+              (input_y - (1 << 10) * y0) *
+              ((1 << 10) - (input_x - (1 << 10) * x0));
+          const int64_t output_20_rl =
+              static_cast<int64_t>(
+                  input_data[Offset(input_shape, b, y0, x1, c)]) *
+              ((1 << 10) - (input_y - (1 << 10) * y0)) *
+              (input_x - (1 << 10) * x0);
+          const int64_t output_20_ru =
+              static_cast<int64_t>(
+                  input_data[Offset(input_shape, b, y1, x1, c)]) *
+              (input_y - (1 << 10) * y0) * (input_x - (1 << 10) * x0);
+          const int64_t output_20 =
+              output_20_ll + output_20_lu + output_20_rl + output_20_ru;
+          const int64_t round = (output_20 > 0) ? (1 << 19) : -(1 << 19);
+          const T interpolation =
+              static_cast<T>((output_20 + round) / (1 << 20));
+          output_data[Offset(output_shape, b, y, x, c)] = interpolation;
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_BILINEAR_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h

@@ -0,0 +1,101 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_RESIZE_NEAREST_NEIGHBOR_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_NEAREST_NEIGHBOR_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline int32_t GetNearestNeighbor(const int input_value,
+                                  const int32_t input_size,
+                                  const int32_t output_size,
+                                  const bool align_corners,
+                                  const bool half_pixel_centers) {
+  const float scale =
+      (align_corners && output_size > 1)
+          ? (input_size - 1) / static_cast<float>(output_size - 1)
+          : input_size / static_cast<float>(output_size);
+  const float offset = half_pixel_centers ? 0.5f : 0.0f;
+  int32_t output_value = std::min(
+      align_corners
+          ? static_cast<int32_t>(TfLiteRound((input_value + offset) * scale))
+          : static_cast<int32_t>(std::floor((input_value + offset) * scale)),
+      input_size - 1);
+  if (half_pixel_centers) {
+    output_value = std::max(static_cast<int32_t>(0), output_value);
+  }
+  return output_value;
+}
+
+template <typename T>
+inline void ResizeNearestNeighbor(
+    const tflite::ResizeNearestNeighborParams& op_params,
+    const RuntimeShape& unextended_input_shape, const T* input_data,
+    const RuntimeShape& output_size_shape, const int32_t* output_size_data,
+    const RuntimeShape& unextended_output_shape, T* output_data) {
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  int32_t batches = MatchingDim(input_shape, 0, output_shape, 0);
+  int32_t input_height = input_shape.Dims(1);
+  int32_t input_width = input_shape.Dims(2);
+  int32_t depth = MatchingDim(input_shape, 3, output_shape, 3);
+
+  // The Tensorflow version of this op allows resize on the width and height
+  // axis only.
+  TFLITE_DCHECK_EQ(output_size_shape.FlatSize(), 2);
+  int32_t output_height = output_size_data[0];
+  int32_t output_width = output_size_data[1];
+
+  const int col_offset = input_shape.Dims(3);
+  const int row_offset = input_shape.Dims(2) * col_offset;
+  const int batch_offset = input_shape.Dims(1) * row_offset;
+
+  const T* input_ptr = input_data;
+  T* output_ptr = output_data;
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < output_height; ++y) {
+      int32_t in_y = GetNearestNeighbor(y, input_height, output_height,
+                                        op_params.align_corners,
+                                        op_params.half_pixel_centers);
+      const T* y_input_ptr = input_ptr + in_y * row_offset;
+      for (int x = 0; x < output_width; ++x) {
+        int32_t in_x = GetNearestNeighbor(x, input_width, output_width,
+                                          op_params.align_corners,
+                                          op_params.half_pixel_centers);
+        const T* x_input_ptr = y_input_ptr + in_x * col_offset;
+        memcpy(output_ptr, x_input_ptr, depth * sizeof(T));
+        output_ptr += depth;
+      }
+    }
+    input_ptr += batch_offset;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_NEAREST_NEIGHBOR_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/round.h

@@ -0,0 +1,51 @@
+/* Copyright 2018 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_KERNELS_INTERNAL_REFERENCE_ROUND_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline float RoundToNearest(float value) {
+  auto floor_val = std::floor(value);
+  auto diff = value - floor_val;
+  if ((diff < 0.5f) ||
+      ((diff == 0.5f) && (static_cast<int>(floor_val) % 2 == 0))) {
+    return floor_val;
+  } else {
+    return floor_val = floor_val + 1.0f;
+  }
+}
+
+inline void Round(const RuntimeShape& input_shape, const float* input_data,
+                  const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    // Note that this implementation matches that of tensorFlow tf.round
+    // and corresponds to the bankers rounding method.
+    // cfenv (for fesetround) is not yet supported universally on Android, so
+    // using a work around.
+    output_data[i] = RoundToNearest(input_data[i]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/slice.h

@@ -0,0 +1,80 @@
+/* Copyright 2021 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_KERNELS_INTERNAL_REFERENCE_SLICE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SLICE_H_
+
+#include "tensorflow/lite/kernels/internal/portable_tensor.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void Slice(const tflite::SliceParams& op_params,
+                  const RuntimeShape& input_shape,
+                  const RuntimeShape& output_shape,
+                  SequentialTensorWriter<T>* writer) {
+  const RuntimeShape ext_shape = RuntimeShape::ExtendedShape(5, input_shape);
+  TFLITE_DCHECK_LE(op_params.begin_count, 5);
+  TFLITE_DCHECK_LE(op_params.size_count, 5);
+  const int begin_count = op_params.begin_count;
+  const int size_count = op_params.size_count;
+  // We front-pad the begin and size vectors.
+  int start[5];
+  int stop[5];
+  for (int i = 0; i < 5; ++i) {
+    int padded_i = 5 - i;
+    start[i] =
+        begin_count < padded_i ? 0 : op_params.begin[begin_count - padded_i];
+    stop[i] =
+        (size_count < padded_i || op_params.size[size_count - padded_i] == -1)
+            ? ext_shape.Dims(i)
+            : start[i] + op_params.size[size_count - padded_i];
+  }
+
+  for (int i0 = start[0]; i0 < stop[0]; ++i0) {
+    for (int i1 = start[1]; i1 < stop[1]; ++i1) {
+      for (int i2 = start[2]; i2 < stop[2]; ++i2) {
+        for (int i3 = start[3]; i3 < stop[3]; ++i3) {
+          for (int i4 = start[4]; i4 < stop[4]; ++i4) {
+            writer->Write(Offset(ext_shape, i0, i1, i2, i3, i4));
+          }
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+inline void Slice(const tflite::SliceParams& op_params,
+                  const RuntimeShape& input_shape, const T* input_data,
+                  const RuntimeShape& output_shape, T* output_data) {
+  SequentialTensorWriter<T> writer(input_data, output_data);
+  return Slice(op_params, input_shape, output_shape, &writer);
+}
+
+template <typename T>
+inline void Slice(const tflite::SliceParams& op_params,
+                  const RuntimeShape& input_shape, const TfLiteTensor* input,
+                  const RuntimeShape& output_shape, TfLiteTensor* output) {
+  SequentialTensorWriter<T> writer(input, output);
+  return Slice(op_params, input_shape, output_shape, &writer);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SLICE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/softmax.h

@@ -0,0 +1,232 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_
+
+#include <limits>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/kernels/op_macros.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void Softmax(const SoftmaxParams& params,
+                    const RuntimeShape& input_shape, const float* input_data,
+                    const RuntimeShape& output_shape, float* output_data) {
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+
+  for (int i = 0; i < outer_size; ++i) {
+    // Find max element value which we'll use to ensure numerical stability
+    // taking advantage of the following equality:
+    // exp(x[i])/sum(exp(x[i])) == exp(x[i]+C)/sum(exp(x[i]+C))
+    float max = std::numeric_limits<float>::lowest();
+    for (int c = 0; c < depth; ++c) {
+      max = std::max(max, input_data[i * depth + c]);
+    }
+
+    // Compute sum.
+    float sum = 0.f;
+    for (int c = 0; c < depth; ++c) {
+      const float exp_c = std::exp((input_data[i * depth + c] - max) *
+                                   static_cast<float>(params.beta));
+      output_data[i * depth + c] = exp_c;
+      sum += exp_c;
+    }
+
+    // Compute result.
+    for (int c = 0; c < depth; ++c) {
+      output_data[i * depth + c] = output_data[i * depth + c] / sum;
+    }
+  }
+}
+
+// Quantized softmax with int8_t/uint8_t input and int8_t/uint8_t/int16_t
+// output.
+template <typename InputT, typename OutputT>
+inline void Softmax(const SoftmaxParams& params,
+                    const RuntimeShape& input_shape, const InputT* input_data,
+                    const RuntimeShape& output_shape, OutputT* output_data) {
+  const int32_t input_beta_multiplier = params.input_multiplier;
+  const int32_t input_beta_left_shift = params.input_left_shift;
+  const int diff_min = params.diff_min;
+  // The representation chosen for the input to the exp() function is Q5.26.
+  // We need to leave extra space since values that we skip might be as large as
+  // -32 before multiplying by input_beta_multiplier, and therefore as large as
+  // -16 afterwards.  Note that exp(-8) is definitely not insignificant to
+  // accumulation, but exp(-16) definitely is.
+  static const int kScaledDiffIntegerBits = 5;
+  static const int kAccumulationIntegerBits = 12;
+  using FixedPointScaledDiff =
+      gemmlowp::FixedPoint<int32_t, kScaledDiffIntegerBits>;
+  using FixedPointAccum =
+      gemmlowp::FixedPoint<int32_t, kAccumulationIntegerBits>;
+  using FixedPoint0 = gemmlowp::FixedPoint<int32_t, 0>;
+
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+
+  for (int i = 0; i < outer_size; ++i) {
+    InputT max_in_row = std::numeric_limits<InputT>::min();
+    for (int c = 0; c < depth; ++c) {
+      max_in_row = std::max(max_in_row, input_data[i * depth + c]);
+    }
+
+    FixedPointAccum sum_of_exps = FixedPointAccum::Zero();
+    for (int c = 0; c < depth; ++c) {
+      int32_t input_diff =
+          static_cast<int32_t>(input_data[i * depth + c]) - max_in_row;
+      if (input_diff >= diff_min) {
+        const int32_t input_diff_rescaled =
+            MultiplyByQuantizedMultiplierGreaterThanOne(
+                input_diff, input_beta_multiplier, input_beta_left_shift);
+        const FixedPointScaledDiff scaled_diff_f8 =
+            FixedPointScaledDiff::FromRaw(input_diff_rescaled);
+        sum_of_exps = sum_of_exps + gemmlowp::Rescale<kAccumulationIntegerBits>(
+                                        exp_on_negative_values(scaled_diff_f8));
+      }
+    }
+
+    int num_bits_over_unit;
+    FixedPoint0 shifted_scale = FixedPoint0::FromRaw(GetReciprocal(
+        sum_of_exps.raw(), kAccumulationIntegerBits, &num_bits_over_unit));
+
+    for (int c = 0; c < depth; ++c) {
+      int32_t input_diff =
+          static_cast<int32_t>(input_data[i * depth + c]) - max_in_row;
+      if (input_diff >= diff_min) {
+        const int32_t input_diff_rescaled =
+            MultiplyByQuantizedMultiplierGreaterThanOne(
+                input_diff, input_beta_multiplier, input_beta_left_shift);
+        const FixedPointScaledDiff scaled_diff_f8 =
+            FixedPointScaledDiff::FromRaw(input_diff_rescaled);
+
+        FixedPoint0 exp_in_0 = exp_on_negative_values(scaled_diff_f8);
+        int32_t unsat_output = gemmlowp::RoundingDivideByPOT(
+            (shifted_scale * exp_in_0).raw(),
+            num_bits_over_unit + 31 - (sizeof(OutputT) * 8));
+
+        const int32_t shifted_output =
+            unsat_output +
+            static_cast<int32_t>(std::numeric_limits<OutputT>::min());
+
+        output_data[i * depth + c] = static_cast<OutputT>(std::max(
+            std::min(shifted_output,
+                     static_cast<int32_t>(std::numeric_limits<OutputT>::max())),
+            static_cast<int32_t>(std::numeric_limits<OutputT>::min())));
+      } else {
+        output_data[i * depth + c] = std::numeric_limits<OutputT>::min();
+      }
+    }
+  }
+}
+
+// Computes exp(input - max_input)
+inline int16_t SoftMaxCalculateExp(const SoftmaxParams& params,
+                                   const int16_t* input_data, const int depth,
+                                   int16_t max_in_row, int i, int c) {
+  int32_t input_diff = input_data[i * depth + c] - max_in_row;
+  // scale the input_diff such that [-65535, 0] correspond to [-10.0, 0.0]
+  // exp lut generated with range [-10, 0], as exp(-10) is negligible.
+  int32_t scaled_diff = MultiplyByQuantizedMultiplier(
+      input_diff, params.input_multiplier, params.input_left_shift);
+  // recenter to [-32768, 32767]
+  int32_t sym_scaled_diff = scaled_diff + 32767;
+  int16_t sat_sym_scaled_diff =
+      std::min(std::max(sym_scaled_diff, static_cast<int32_t>(-32768)),
+               static_cast<int32_t>(32767));
+  // apply the exp() LUT activation function
+  return lut_lookup(sat_sym_scaled_diff, params.exp_lut);
+}
+// Quantized softmax with int16_t input and int16_t output.
+inline void SoftmaxInt16(const SoftmaxParams& params,
+                         const RuntimeShape& input_shape,
+                         const int16_t* input_data,
+                         const RuntimeShape& output_shape,
+                         int16_t* output_data) {
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+
+  for (int i = 0; i < outer_size; ++i) {
+    // Find the largest element
+    int16_t max_in_row = std::numeric_limits<int16_t>::min();
+    for (int c = 0; c < depth; ++c) {
+      max_in_row = std::max(max_in_row, input_data[i * depth + c]);
+    }
+
+    // This loops computes the exp values and their sum. We will need the exp
+    // values later on in the function so we cache them in the output_data
+    // buffer. This is an optimization done to avoid calculating the exp values
+    // twice making use of the output_data buffer as scratch memory.
+    int32_t sum_of_exps = 0;  // Q16.15 fixed point format.
+    int16_t* exp_results_Q015 = output_data + i * depth;
+    for (int c = 0; c < depth; ++c) {
+      exp_results_Q015[c] =
+          SoftMaxCalculateExp(params, input_data, depth, max_in_row, i, c);
+      sum_of_exps += exp_results_Q015[c];
+    }
+
+    // Compute the reciprocal 1/sum_of_exps
+    uint8_t headroom_plus_one =
+        CountLeadingZeros(static_cast<uint32_t>(sum_of_exps));
+    int32_t shifted_sum =
+        ((static_cast<int64_t>(sum_of_exps) << (headroom_plus_one - 1)) +
+         (1 << 13)) >>
+        14;
+    // since the LUT computes 1/(1 + x) we need to first compute x = (sum - 1).
+    // also, the LUT expects a symmetrical input, so we must also recenter x
+    // from [0, 65535] to [-32768, 32767].
+    int32_t sym_shifted_sum = shifted_sum + (-((1 << 15) + (1 << 16)));
+    int16_t sat_sym_shifted_sum = static_cast<int16_t>(
+        std::min(std::max(sym_shifted_sum, static_cast<int32_t>(-32768)),
+                 static_cast<int32_t>(32767)));
+    // apply 1/(1 + x) LUT activation function
+    int16_t reciprocal_scale_Q015 =
+        lut_lookup(sat_sym_shifted_sum, params.one_over_one_plus_x_lut);
+
+    // Rescale the exp_result with reciprocal
+    // range of output is [0, 32767] correspond to [0.0, 1.0]
+    for (int c = 0; c < depth; ++c) {
+      uint8_t right_shift = 31 - headroom_plus_one;
+      int64_t round = 1 << (right_shift - 1);
+      int32_t result = (static_cast<int64_t>(exp_results_Q015[c]) *
+                            static_cast<int64_t>(reciprocal_scale_Q015) +
+                        round) >>
+                       right_shift;
+      output_data[i * depth + c] = static_cast<int16_t>(
+          std::min(std::max(result, static_cast<int32_t>(0)),
+                   static_cast<int32_t>(32767)));
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SOFTMAX_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/space_to_batch_nd.h

@@ -0,0 +1,109 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_SPACE_TO_BATCH_ND_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SPACE_TO_BATCH_ND_H_
+
+#include <cmath>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+// TODO(b/135760455): Move this method anonymous namespace in a cc file.
+inline RuntimeShape ExtendShapeSpaceToBatch(const RuntimeShape& shape) {
+  if (shape.DimensionsCount() == 4) {
+    return shape;
+  }
+  RuntimeShape new_shape(4, 1);
+  new_shape.SetDim(0, shape.Dims(0));
+  new_shape.SetDim(1, shape.Dims(1));
+  new_shape.SetDim(3, shape.Dims(2));
+  return new_shape;
+}
+
+template <typename T>
+inline void SpaceToBatchND(const SpaceToBatchParams& params,
+                           const RuntimeShape& unextended_input1_shape,
+                           const T* input1_data,
+                           const RuntimeShape& unextended_input2_shape,
+                           const int32_t* block_shape_data,
+                           const RuntimeShape& unextended_input3_shape,
+                           const int32_t* paddings_data,
+                           const RuntimeShape& unextended_output_shape,
+                           T* output_data) {
+  ruy::profiler::ScopeLabel label("SpaceToBatchND");
+  TFLITE_DCHECK_GE(unextended_input1_shape.DimensionsCount(), 3);
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(unextended_input1_shape.DimensionsCount(),
+                   unextended_output_shape.DimensionsCount());
+
+  // Extends the input/output shape from 3D to 4D if needed, NHC -> NH1C.
+  const RuntimeShape input1_shape =
+      ExtendShapeSpaceToBatch(unextended_input1_shape);
+  const RuntimeShape output_shape =
+      ExtendShapeSpaceToBatch(unextended_output_shape);
+
+  const int depth = input1_shape.Dims(3);
+  const int input_width = input1_shape.Dims(2);
+  const int input_height = input1_shape.Dims(1);
+  const int input_batch_size = input1_shape.Dims(0);
+
+  const int output_width = output_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_batch_size = output_shape.Dims(0);
+
+  const int block_shape_height = block_shape_data[0];
+  const int block_shape_width =
+      unextended_input1_shape.DimensionsCount() == 4 ? block_shape_data[1] : 1;
+  const int padding_top = paddings_data[0];
+  const int padding_left =
+      unextended_input1_shape.DimensionsCount() == 4 ? paddings_data[2] : 0;
+
+  // For uint8 quantized, the correct padding "zero value" is the output offset.
+  const int32_t pad_value = params.output_offset;
+  for (int out_b = 0; out_b < output_batch_size; ++out_b) {
+    int input_batch = out_b % input_batch_size;
+    int shift_w = (out_b / input_batch_size) % block_shape_width;
+    int shift_h = (out_b / input_batch_size) / block_shape_width;
+    for (int out_h = 0; out_h < output_height; ++out_h) {
+      for (int out_w = 0; out_w < output_width; ++out_w) {
+        T* out = output_data + Offset(output_shape, out_b, out_h, out_w, 0);
+        if (out_h * block_shape_height + shift_h < padding_top ||
+            out_h * block_shape_height + shift_h >=
+                padding_top + input_height ||
+            out_w * block_shape_width + shift_w < padding_left ||
+            out_w * block_shape_width + shift_w >= padding_left + input_width) {
+          // This may not execute correctly when pad_value != 0 and T != uint8.
+          memset(out, pad_value, depth * sizeof(T));
+        } else {
+          const T* in =
+              input1_data +
+              Offset(input1_shape, input_batch,
+                     (out_h * block_shape_height + shift_h) - padding_top,
+                     (out_w * block_shape_width + shift_w) - padding_left, 0);
+          memcpy(out, in, depth * sizeof(T));
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SPACE_TO_BATCH_ND_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/space_to_depth.h

@@ -0,0 +1,80 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_SPACE_TO_DEPTH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SPACE_TO_DEPTH_H_
+
+#include <cstdint>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+inline void SpaceToDepth(const tflite::SpaceToDepthParams& op_params,
+                         const RuntimeShape& unextended_input_shape,
+                         const T* input_data,
+                         const RuntimeShape& unextended_output_shape,
+                         T* output_data) {
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  const int input_depth = input_shape.Dims(3);
+  const int input_width = input_shape.Dims(2);
+  const int input_height = input_shape.Dims(1);
+  const int input_batch = input_shape.Dims(0);
+
+  const int output_depth = output_shape.Dims(3);
+  const int output_width = output_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_batch = output_shape.Dims(0);
+
+  const int32_t block_size = op_params.block_size;
+
+  TFLITE_DCHECK_EQ(input_width, output_width * block_size);
+  TFLITE_DCHECK_EQ(input_height, output_height * block_size);
+  TFLITE_DCHECK_EQ(input_depth * block_size * block_size, output_depth);
+  TFLITE_DCHECK_EQ(input_batch, output_batch);
+
+  for (int in_b = 0; in_b < input_batch; ++in_b) {
+    for (int in_h = 0; in_h < input_height; ++in_h) {
+      for (int in_w = 0; in_w < input_width; ++in_w) {
+        for (int in_d = 0; in_d < input_depth; ++in_d) {
+          const int out_d =
+              in_d + ((in_h % block_size) * block_size + in_w % block_size) *
+                         input_depth;
+          const int out_w = in_w / block_size;
+          const int out_h = in_h / block_size;
+          const int out_b = in_b;
+
+          const int input_index = Offset(input_shape, in_b, in_h, in_w, in_d);
+          const int output_index =
+              Offset(output_shape, out_b, out_h, out_w, out_d);
+
+          output_data[output_index] = input_data[input_index];
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SPACE_TO_DEPTH_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/strided_slice.h

@@ -0,0 +1,121 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/portable_tensor.h"
+#include "tensorflow/lite/kernels/internal/strided_slice_logic.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void StridedSlice(const tflite::StridedSliceParams& op_params,
+                         const RuntimeShape& unextended_input_shape,
+                         const RuntimeShape& unextended_output_shape,
+                         SequentialTensorWriter<T>* writer) {
+  using strided_slice::LoopCondition;
+  using strided_slice::StartForAxis;
+  using strided_slice::StopForAxis;
+
+  ruy::profiler::ScopeLabel label("StridedSlice");
+
+  // Note that the output_shape is not used herein.
+  tflite::StridedSliceParams params_copy = op_params;
+
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 5);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(5, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(5, unextended_output_shape);
+
+  // Reverse and pad to 5 dimensions because that is what the runtime code
+  // requires (ie. all shapes must be 5D and are given backwards).
+  strided_slice::StridedSlicePadIndices(&params_copy, 5);
+
+  const int start_0 = StartForAxis(params_copy, input_shape, 0);
+  const int stop_0 = StopForAxis(params_copy, input_shape, 0, start_0);
+  const int start_1 = StartForAxis(params_copy, input_shape, 1);
+  const int stop_1 = StopForAxis(params_copy, input_shape, 1, start_1);
+  const int start_2 = StartForAxis(params_copy, input_shape, 2);
+  const int stop_2 = StopForAxis(params_copy, input_shape, 2, start_2);
+  const int start_3 = StartForAxis(params_copy, input_shape, 3);
+  const int stop_3 = StopForAxis(params_copy, input_shape, 3, start_3);
+  const int start_4 = StartForAxis(params_copy, input_shape, 4);
+  const int stop_4 = StopForAxis(params_copy, input_shape, 4, start_4);
+
+  for (int offset_0 = start_0 * input_shape.Dims(1),
+           end_0 = stop_0 * input_shape.Dims(1),
+           step_0 = params_copy.strides[0] * input_shape.Dims(1);
+       !LoopCondition(offset_0, end_0, params_copy.strides[0]);
+       offset_0 += step_0) {
+    for (int offset_1 = (offset_0 + start_1) * input_shape.Dims(2),
+             end_1 = (offset_0 + stop_1) * input_shape.Dims(2),
+             step_1 = params_copy.strides[1] * input_shape.Dims(2);
+         !LoopCondition(offset_1, end_1, params_copy.strides[1]);
+         offset_1 += step_1) {
+      for (int offset_2 = (offset_1 + start_2) * input_shape.Dims(3),
+               end_2 = (offset_1 + stop_2) * input_shape.Dims(3),
+               step_2 = params_copy.strides[2] * input_shape.Dims(3);
+           !LoopCondition(offset_2, end_2, params_copy.strides[2]);
+           offset_2 += step_2) {
+        for (int offset_3 = (offset_2 + start_3) * input_shape.Dims(4),
+                 end_3 = (offset_2 + stop_3) * input_shape.Dims(4),
+                 step_3 = params_copy.strides[3] * input_shape.Dims(4);
+             !LoopCondition(offset_3, end_3, params_copy.strides[3]);
+             offset_3 += step_3) {
+          for (int offset_4 = offset_3 + start_4, end_4 = offset_3 + stop_4;
+               !LoopCondition(offset_4, end_4, params_copy.strides[4]);
+               offset_4 += params_copy.strides[4]) {
+            writer->Write(offset_4);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+inline void StridedSlice(const tflite::StridedSliceParams& op_params,
+                         const RuntimeShape& unextended_input_shape,
+                         const T* input_data,
+                         const RuntimeShape& unextended_output_shape,
+                         T* output_data) {
+  SequentialTensorWriter<T> writer(input_data, output_data);
+  StridedSlice<T>(op_params, unextended_input_shape, unextended_output_shape,
+                  &writer);
+}
+
+template <typename T>
+inline void StridedSlice(const tflite::StridedSliceParams& op_params,
+                         const RuntimeShape& unextended_input_shape,
+                         const TfLiteTensor* input,
+                         const RuntimeShape& unextended_output_shape,
+                         TfLiteTensor* output) {
+  SequentialTensorWriter<T> writer(input, output);
+  StridedSlice<T>(op_params, unextended_input_shape, unextended_output_shape,
+                  &writer);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRIDED_SLICE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/sub.h

@@ -0,0 +1,476 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_SUB_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SUB_H_
+
+#include <stdint.h>
+
+#include <algorithm>
+#include <limits>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void SubNonBroadcast(const ArithmeticParams& params,
+                            const RuntimeShape& input1_shape,
+                            const float* input1_data,
+                            const RuntimeShape& input2_shape,
+                            const float* input2_data,
+                            const RuntimeShape& output_shape,
+                            float* output_data) {
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] - input2_data[i], params.float_activation_min,
+        params.float_activation_max);
+  }
+}
+
+inline void SubNonBroadcast(const ArithmeticParams& params,
+                            const RuntimeShape& input1_shape,
+                            const int32_t* input1_data,
+                            const RuntimeShape& input2_shape,
+                            const int32_t* input2_data,
+                            const RuntimeShape& output_shape,
+                            int32_t* output_data) {
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] - input2_data[i], params.quantized_activation_min,
+        params.quantized_activation_max);
+  }
+}
+
+// TODO(b/151345304): We can implement BroadcastSub on buffers of arbitrary
+// dimensionality if the runtime code does a single loop over one dimension
+// that handles broadcasting as the base case. The code generator would then
+// generate max(D1, D2) nested for loops.
+template <int N = 5>
+inline void BroadcastSubSlow(const ArithmeticParams& params,
+                             const RuntimeShape& input1_shape,
+                             const float* input1_data,
+                             const RuntimeShape& input2_shape,
+                             const float* input2_data,
+                             const RuntimeShape& output_shape,
+                             float* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastSubSlow/float");
+  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
+  NdArrayDesc<N> desc1;
+  NdArrayDesc<N> desc2;
+  NdArrayDesc<N> output_desc;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  auto sub_func = [&](int indexes[N]) {
+    output_data[SubscriptToIndex(output_desc, indexes)] =
+        ActivationFunctionWithMinMax(
+            input1_data[SubscriptToIndex(desc1, indexes)] -
+                input2_data[SubscriptToIndex(desc2, indexes)],
+            params.float_activation_min, params.float_activation_max);
+  };
+  NDOpsHelper<N>(output_desc, sub_func);
+}
+
+template <int N = 5>
+inline void BroadcastSubSlow(const ArithmeticParams& params,
+                             const RuntimeShape& input1_shape,
+                             const int32_t* input1_data,
+                             const RuntimeShape& input2_shape,
+                             const int32_t* input2_data,
+                             const RuntimeShape& output_shape,
+                             int32_t* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastSubSlow/int32_t");
+  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
+  NdArrayDesc<N> desc1;
+  NdArrayDesc<N> desc2;
+  NdArrayDesc<N> output_desc;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  auto sub_func = [&](int indexes[N]) {
+    output_data[SubscriptToIndex(output_desc, indexes)] =
+        ActivationFunctionWithMinMax(
+            input1_data[SubscriptToIndex(desc1, indexes)] -
+                input2_data[SubscriptToIndex(desc2, indexes)],
+            params.quantized_activation_min, params.quantized_activation_max);
+  };
+  NDOpsHelper<N>(output_desc, sub_func);
+}
+
+template <int N = 5>
+void BroadcastSubSlow(const ArithmeticParams& params,
+                      const RuntimeShape& input1_shape,
+                      const int64_t* input1_data,
+                      const RuntimeShape& input2_shape,
+                      const int64_t* input2_data,
+                      const RuntimeShape& output_shape, int64_t* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastSubSlow/int64_t");
+  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
+  NdArrayDesc<N> desc1;
+  NdArrayDesc<N> desc2;
+  NdArrayDesc<N> output_desc;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  auto sub_func = [&](int indexes[N]) {
+    output_data[SubscriptToIndex(output_desc, indexes)] =
+        ActivationFunctionWithMinMax(
+            input1_data[SubscriptToIndex(desc1, indexes)] -
+                input2_data[SubscriptToIndex(desc2, indexes)],
+            params.int64_activation_min, params.int64_activation_max);
+  };
+  NDOpsHelper<N>(output_desc, sub_func);
+}
+
+template <typename T, int N = 5>
+void BroadcastSubSlow(const ArithmeticParams& params,
+                      const RuntimeShape& input1_shape, const T* input1_data,
+                      const RuntimeShape& input2_shape, const T* input2_data,
+                      const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastSubSlow/templated");
+  TFLITE_DCHECK_LE(input1_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(input2_shape.DimensionsCount(), N);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), N);
+  NdArrayDesc<N> desc1;
+  NdArrayDesc<N> desc2;
+  NdArrayDesc<N> output_desc;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  auto sub_func = [&](int indexes[N]) {
+    output_data[SubscriptToIndex(output_desc, indexes)] =
+        ActivationFunctionWithMinMax(
+            input1_data[SubscriptToIndex(desc1, indexes)] -
+                input2_data[SubscriptToIndex(desc2, indexes)],
+            params.quantized_activation_min, params.quantized_activation_max);
+  };
+  NDOpsHelper<N>(output_desc, sub_func);
+}
+
+template <int N = 5>
+inline void BroadcastSub16POTSlow(const ArithmeticParams& params,
+                                  const RuntimeShape& input1_shape,
+                                  const int16_t* input1_data,
+                                  const RuntimeShape& input2_shape,
+                                  const int16_t* input2_data,
+                                  const RuntimeShape& output_shape,
+                                  int16_t* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastSub16POTSlow/int16_t");
+  NdArrayDesc<N> desc1;
+  NdArrayDesc<N> desc2;
+  NdArrayDesc<N> output_desc;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  auto sub_func = [&](int indexes[N]) {
+    const int32_t input1_val = input1_data[SubscriptToIndex(desc1, indexes)];
+    const int32_t input2_val = input2_data[SubscriptToIndex(desc2, indexes)];
+    const int32_t scaled_input1_val =
+        gemmlowp::RoundingDivideByPOT(input1_val, -params.input1_shift);
+    const int32_t scaled_input2_val =
+        gemmlowp::RoundingDivideByPOT(input2_val, -params.input2_shift);
+    const int32_t raw_output = scaled_input1_val - scaled_input2_val;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[SubscriptToIndex(output_desc, indexes)] =
+        static_cast<int16_t>(clamped_output);
+  };
+  NDOpsHelper<N>(output_desc, sub_func);
+}
+
+template <typename T, int N = 5>
+void BroadcastQuantSubSlow(const ArithmeticParams& params,
+                           const RuntimeShape& input1_shape,
+                           const T* input1_data,
+                           const RuntimeShape& input2_shape,
+                           const T* input2_data,
+                           const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastQuantSubSlow/T");
+  NdArrayDesc<N> desc1;
+  NdArrayDesc<N> desc2;
+  NdArrayDesc<N> output_desc;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, output_shape), &output_desc);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  auto sub_func = [&](int indexes[N]) {
+    const int32_t input1_val =
+        params.input1_offset + input1_data[SubscriptToIndex(desc1, indexes)];
+    const int32_t input2_val =
+        params.input2_offset + input2_data[SubscriptToIndex(desc2, indexes)];
+    const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+    const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+    const int32_t scaled_input1_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input1_val, params.input1_multiplier, params.input1_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, params.input2_multiplier, params.input2_shift);
+    const int32_t raw_sub = scaled_input1_val - scaled_input2_val;
+    const int32_t raw_output =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            raw_sub, params.output_multiplier, params.output_shift) +
+        params.output_offset;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[SubscriptToIndex(output_desc, indexes)] =
+        static_cast<T>(clamped_output);
+  };
+  NDOpsHelper<N>(output_desc, sub_func);
+}
+
+// Element-wise add that can often be used for inner loop of broadcast add as
+// well as the non-broadcast add.
+template <typename T>
+inline void SubElementwise(int size, const ArithmeticParams& params,
+                           const T* input1_data, const T* input2_data,
+                           T* output_data) {
+  for (int i = 0; i < size; ++i) {
+    const int32_t input1_val = params.input1_offset + input1_data[i];
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+    const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+    const int32_t scaled_input1_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input1_val, params.input1_multiplier, params.input1_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, params.input2_multiplier, params.input2_shift);
+    const int32_t raw_sub = scaled_input1_val - scaled_input2_val;
+    const int32_t raw_output =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            raw_sub, params.output_multiplier, params.output_shift) +
+        params.output_offset;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+inline void Sub(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const uint8_t* input1_data,
+                const RuntimeShape& input2_shape, const uint8_t* input2_data,
+                const RuntimeShape& output_shape, uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  TFLITE_DCHECK_GT(params.input1_offset, -256);
+  TFLITE_DCHECK_GT(params.input2_offset, -256);
+  TFLITE_DCHECK_LT(params.input1_offset, 256);
+  TFLITE_DCHECK_LT(params.input2_offset, 256);
+  SubElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void Sub(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int8_t* input1_data,
+                const RuntimeShape& input2_shape, const int8_t* input2_data,
+                const RuntimeShape& output_shape, int8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  TFLITE_DCHECK_GE(params.input1_offset, -128);
+  TFLITE_DCHECK_GE(params.input2_offset, -128);
+  // offset = -quantization_params.zero_point in PrepareGeneralSubOp().
+  // So it's maximum can be 128 not 127.
+  TFLITE_DCHECK_LE(params.input1_offset, 128);
+  TFLITE_DCHECK_LE(params.input2_offset, 128);
+  SubElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void Sub(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int16_t* input1_data,
+                const RuntimeShape& input2_shape, const int16_t* input2_data,
+                const RuntimeShape& output_shape, int16_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  TFLITE_DCHECK_EQ(params.input1_offset, 0);
+  TFLITE_DCHECK_EQ(params.input2_offset, 0);
+  SubElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+template <typename T>
+void Sub(const ArithmeticParams& params, const RuntimeShape& input1_shape,
+         const T* input1_data, const RuntimeShape& input2_shape,
+         const T* input2_data, const RuntimeShape& output_shape,
+         T* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)] -
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+        }
+      }
+    }
+  }
+}
+
+inline void SetActivationMinMax(const ArithmeticParams& params,
+                                int32_t* activation_min,
+                                int32_t* activation_max) {
+  *activation_min = params.quantized_activation_min;
+  *activation_max = params.quantized_activation_max;
+}
+
+inline void SetActivationMinMax(const ArithmeticParams& params,
+                                float* activation_min, float* activation_max) {
+  *activation_min = params.float_activation_min;
+  *activation_max = params.float_activation_max;
+}
+
+inline void SetActivationMinMax(const ArithmeticParams& params,
+                                int64_t* activation_min,
+                                int64_t* activation_max) {
+  *activation_min = params.int64_activation_min;
+  *activation_max = params.int64_activation_max;
+}
+
+template <typename T>
+inline void SubWithActivation(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("SubWithActivation");
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  T activation_min, activation_max;
+  SetActivationMinMax(params, &activation_min, &activation_max);
+
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] - input2_data[i], activation_min, activation_max);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SUB_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/tanh.h

@@ -0,0 +1,129 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_TANH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TANH_H_
+
+#include <cmath>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/kernels/op_macros.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void Tanh(const RuntimeShape& input_shape, const float* input_data,
+                 const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    float val = input_data[i];
+    float result = std::tanh(val);
+    output_data[i] = result;
+  }
+}
+
+// Convenience version that allows, for example, generated-code calls to be
+// uniform between data types.
+inline void Tanh(const TanhParams&, const RuntimeShape& input_shape,
+                 const float* input_data, const RuntimeShape& output_shape,
+                 float* output_data) {
+  // Drop params: not needed.
+  Tanh(input_shape, input_data, output_shape, output_data);
+}
+
+inline void Tanh(const TanhParams& params, const RuntimeShape& input_shape,
+                 const int16_t* input_data, const RuntimeShape& output_shape,
+                 int16_t* output_data) {
+  const int input_left_shift = params.input_left_shift;
+  // Support for shifts is limited until we have a parameterized version of
+  // SaturatingRoundingMultiplyByPOT().
+  TFLITE_DCHECK_GE(input_left_shift, 0);
+  TFLITE_DCHECK_LE(input_left_shift, 1);
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  // F0 uses 0 integer bits, range [-1, 1].
+  // This is the return type of math functions such as tanh, logistic,
+  // whose range is in [-1, 1].
+  using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+  // F3 uses 3 integer bits, range [-8, 8], the input range expected here.
+  using F3 = gemmlowp::FixedPoint<std::int16_t, 3>;
+
+  if (input_left_shift == 0) {
+    for (int i = 0; i < flat_size; i++) {
+      F3 input = F3::FromRaw(input_data[i]);
+      F0 output = gemmlowp::tanh(input);
+      output_data[i] = output.raw();
+    }
+  } else {
+    for (int i = 0; i < flat_size; i++) {
+      F3 input = F3::FromRaw(
+          gemmlowp::SaturatingRoundingMultiplyByPOT<1>(input_data[i]));
+      F0 output = gemmlowp::tanh(input);
+      output_data[i] = output.raw();
+    }
+  }
+}
+
+inline void Tanh(const TanhParams& params, const RuntimeShape& input_shape,
+                 const uint8_t* input_data, const RuntimeShape& output_shape,
+                 uint8_t* output_data) {
+  const int32_t input_zero_point = params.input_zero_point;
+  const int32_t input_range_radius = params.input_range_radius;
+  const int32_t input_multiplier = params.input_multiplier;
+  const int input_left_shift = params.input_left_shift;
+  const int32_t output_zero_point = 128;
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    const uint8_t input_val_u8 = input_data[i];
+    const int32_t input_val_centered =
+        static_cast<int32_t>(input_val_u8) - input_zero_point;
+    uint8_t output_val;
+    if (input_val_centered <= -input_range_radius) {
+      output_val = 0;
+    } else if (input_val_centered >= input_range_radius) {
+      output_val = 255;
+    } else {
+      const int32_t input_val_rescaled =
+          MultiplyByQuantizedMultiplierGreaterThanOne(
+              input_val_centered, input_multiplier, input_left_shift);
+      using FixedPoint4 = gemmlowp::FixedPoint<int32_t, 4>;
+      using FixedPoint0 = gemmlowp::FixedPoint<int32_t, 0>;
+      const FixedPoint4 input_val_f4 = FixedPoint4::FromRaw(input_val_rescaled);
+      const FixedPoint0 output_val_f0 = gemmlowp::tanh(input_val_f4);
+      // Convert from Q0.31 to Q24.7.
+      using gemmlowp::RoundingDivideByPOT;
+      int32_t output_val_s32 = RoundingDivideByPOT(output_val_f0.raw(), 24);
+      output_val_s32 += output_zero_point;
+      if (output_val_s32 == 256) {
+        output_val_s32 = 255;
+      }
+      // Reinterpret as Q0.7, encoded in uint8_t.
+      TFLITE_DCHECK_GE(output_val_s32, 0);
+      TFLITE_DCHECK_LE(output_val_s32, 255);
+      output_val = static_cast<uint8_t>(output_val_s32);
+    }
+    output_data[i] = output_val;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TANH_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/transpose.h

@@ -0,0 +1,111 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T, int N>
+void TransposeImpl(const TransposeParams& params,
+                   const RuntimeShape& unextended_input_shape,
+                   const T* input_data,
+                   const RuntimeShape& unextended_output_shape,
+                   T* output_data) {
+  const int unextended_input_size = unextended_input_shape.DimensionsCount();
+  const int unextended_output_size = unextended_output_shape.DimensionsCount();
+  TFLITE_DCHECK_LE(unextended_input_size, N);
+  TFLITE_DCHECK_LE(unextended_output_size, N);
+  TFLITE_DCHECK_EQ(unextended_output_size, params.perm_count);
+  const int input_ext_size = N - unextended_input_size;
+  const int output_ext_size = N - unextended_output_size;
+  NdArrayDesc<N> input_desc;
+  NdArrayDesc<N> output_desc;
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, unextended_input_shape),
+                 &input_desc);
+  CopyDimsToDesc(RuntimeShape::ExtendedShape(N, unextended_output_shape),
+                 &output_desc);
+
+  // The perm data is extended to match the output, each index incremented by
+  // the amount of front padding of the input shape.
+  int extended_perm[N];
+  for (int i = 0; i < N; ++i) {
+    extended_perm[i] = i < output_ext_size
+                           ? i
+                           : params.perm[i - output_ext_size] + input_ext_size;
+  }
+
+  // Permutes the input shape so we don't need to permute the indexes inside
+  // the loop. Check to make sure output_dims is matching input_dims.
+  NdArrayDesc<N> perm_input_desc;
+  for (int k = 0; k < N; ++k) {
+    TFLITE_DCHECK_EQ(input_desc.extents[extended_perm[k]],
+                     output_desc.extents[k]);
+    perm_input_desc.extents[k] = input_desc.extents[extended_perm[k]];
+    perm_input_desc.strides[k] = input_desc.strides[extended_perm[k]];
+  }
+
+  // Naive transpose loop (iterate on output index and compute input index).
+  auto tranpose_func = [&](int indexes[N]) {
+    output_data[SubscriptToIndex(output_desc, indexes)] =
+        input_data[SubscriptToIndex(perm_input_desc, indexes)];
+  };
+  NDOpsHelper<N>(output_desc, tranpose_func);
+}
+
+template <typename T, int N = 5>
+void Transpose(const TransposeParams& params,
+               const RuntimeShape& unextended_input_shape, const T* input_data,
+               const RuntimeShape& unextended_output_shape, T* output_data) {
+  // Transpose kernel only does rearranging values not numeric evaluations on
+  // each cell. It's safe to implement per size of scalar type and this trick
+  // keeps the total code size in a reasonable range.
+  switch (sizeof(T)) {
+    case 1:
+      TransposeImpl<int8_t, N>(params, unextended_input_shape,
+                               reinterpret_cast<const int8_t*>(input_data),
+                               unextended_output_shape,
+                               reinterpret_cast<int8_t*>(output_data));
+      break;
+    case 2:
+      TransposeImpl<int16_t, N>(params, unextended_input_shape,
+                                reinterpret_cast<const int16_t*>(input_data),
+                                unextended_output_shape,
+                                reinterpret_cast<int16_t*>(output_data));
+      break;
+
+    case 4:
+      TransposeImpl<int32_t, N>(params, unextended_input_shape,
+                                reinterpret_cast<const int32_t*>(input_data),
+                                unextended_output_shape,
+                                reinterpret_cast<int32_t*>(output_data));
+      break;
+    case 8:
+      TransposeImpl<int64_t, N>(params, unextended_input_shape,
+                                reinterpret_cast<const int64_t*>(input_data),
+                                unextended_output_shape,
+                                reinterpret_cast<int64_t*>(output_data));
+      break;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/transpose_conv.h

@@ -0,0 +1,217 @@
+/* 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_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void TransposeConv(
+    const ConvParams& params, const RuntimeShape& input_shape,
+    const float* input_data, const RuntimeShape& filter_shape,
+    const float* filter_data, const RuntimeShape& bias_shape,
+    const float* bias_data, const RuntimeShape& output_shape,
+    float* output_data, const RuntimeShape& im2col_shape, float* im2col_data) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  // Although transpose convolution simplifies to convolution with transposed
+  // weights for strides of 1, non-unitary striding complicates matters. To
+  // keep this reference implementation as clear as possible, we use a
+  // "scatter" access pattern, where we loop through all the input elements,
+  // computing their influence on the output, rather than looping through the
+  // output elements in the typical "gather" access pattern of a conv. We
+  // therefore must initialize the output array to zero.
+  const int num_elements = output_shape.FlatSize();
+  for (int i = 0; i < num_elements; i++) {
+    output_data[i] = 0.0f;
+  }
+
+  // Loop through input elements one at a time.
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int in_y = 0; in_y < input_height; ++in_y) {
+      for (int in_x = 0; in_x < input_width; ++in_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          // Loop through the output elements it will influence
+          const int out_x_origin = (in_x * stride_width) - pad_width;
+          const int out_y_origin = (in_y * stride_height) - pad_height;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int out_channel = 0; out_channel < output_depth;
+                   ++out_channel) {
+                // Compute output element location
+                const int out_x = out_x_origin + filter_x;
+                const int out_y = out_y_origin + filter_y;
+                // We cannot accumulate out of bounds
+                if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) &&
+                    (out_y < output_height)) {
+                  float input_value = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  float filter_value =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  output_data[Offset(output_shape, batch, out_y, out_x,
+                                     out_channel)] +=
+                      input_value * filter_value;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  if (bias_data) {
+    for (int batch = 0; batch < batches; ++batch) {
+      for (int out_y = 0; out_y < output_height; ++out_y) {
+        for (int out_x = 0; out_x < output_width; ++out_x) {
+          for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+            output_data[Offset(output_shape, batch, out_y, out_x,
+                               out_channel)] += bias_data[out_channel];
+          }
+        }
+      }
+    }
+  }
+}
+
+inline void TransposeConv(
+    const ConvParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& filter_shape,
+    const uint8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    uint8_t* output_data, const RuntimeShape& im2col_shape,
+    uint8_t* im2col_data, int32_t* scratch_buffer) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  const int num_elements = output_shape.FlatSize();
+  // We need to initialize scratch_buffer to all 0s, as we apply the same
+  // 'scatter' based trick as in float version.
+  memset(scratch_buffer, 0, num_elements * sizeof(int32_t));
+
+  // Loop through input elements one at a time.
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int in_y = 0; in_y < input_height; ++in_y) {
+      for (int in_x = 0; in_x < input_width; ++in_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          // Loop through the output elements it will influence.
+          const int out_x_origin = (in_x * stride_width) - pad_width;
+          const int out_y_origin = (in_y * stride_height) - pad_height;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int out_channel = 0; out_channel < output_depth;
+                   ++out_channel) {
+                // Compute output element location.
+                const int out_x = out_x_origin + filter_x;
+                const int out_y = out_y_origin + filter_y;
+                // We cannot accumulate out of bounds.
+                if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) &&
+                    (out_y < output_height)) {
+                  uint8_t input_value = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  uint8_t filter_value =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                        out_channel)] +=
+                      (input_value + input_offset) *
+                      (filter_value + filter_offset);
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          int32_t acc = scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                              out_channel)];
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier, output_shift);
+          scaled_acc += output_offset;
+          scaled_acc = std::max(scaled_acc, output_activation_min);
+          scaled_acc = std::min(scaled_acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<uint8_t>(scaled_acc);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_CONV_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/runtime_shape.h

@@ -0,0 +1,155 @@
+/* Copyright 2021 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_KERNELS_INTERNAL_RUNTIME_SHAPE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_RUNTIME_SHAPE_H_
+
+namespace tflite {
+
+template <int N>
+struct Dims {
+  int sizes[N];
+  int strides[N];
+};
+
+class RuntimeShape {
+ public:
+  RuntimeShape& operator=(RuntimeShape const&) = delete;
+
+  RuntimeShape() : size_(0) {}
+
+  explicit RuntimeShape(int dimensions_count) : size_(dimensions_count) {}
+
+  RuntimeShape(int shape_size, int32_t value) : size_(shape_size) {
+    for (int i = 0; i < shape_size; ++i) {
+      SetDim(i, value);
+    }
+  }
+
+  RuntimeShape(int dimensions_count, const int32_t* dims_data)
+      : size_(dimensions_count) {
+    ReplaceWith(dimensions_count, dims_data);
+  }
+
+  bool operator==(const RuntimeShape& comp) const {
+    return this->size_ == comp.size_ &&
+           std::memcmp(DimsData(), comp.DimsData(), size_ * sizeof(int32_t)) ==
+               0;
+  }
+
+  ~RuntimeShape() {}
+
+  int32_t DimensionsCount() const { return size_; }
+  int32_t Dims(int i) const {
+    TFLITE_DCHECK_GE(i, 0);
+    TFLITE_DCHECK_LT(i, size_);
+    return dims_[i];
+  }
+  void SetDim(int i, int32_t val) {
+    TFLITE_DCHECK_GE(i, 0);
+    TFLITE_DCHECK_LT(i, size_);
+    dims_[i] = val;
+  }
+
+  static RuntimeShape ExtendedShape(int new_shape_size,
+                                    const RuntimeShape& shape) {
+    return RuntimeShape(new_shape_size, shape, 1);
+  }
+  int32_t* DimsData() { return dims_; }
+  const int32_t* DimsData() const { return dims_; }
+  const int32_t* DimsDataUpTo5D() const { return dims_; }
+
+  void ReplaceWith(int dimensions_count, const int32_t* dims_data) {
+    size_ = dimensions_count;
+    int32_t* dst_dims = DimsData();
+    std::memcpy(dst_dims, dims_data, dimensions_count * sizeof(int32_t));
+  }
+
+  // Returns the total count of elements, that is the size when flattened into a
+  // vector.
+  int FlatSize() const {
+    int buffer_size = 1;
+    const int* dims_data = reinterpret_cast<const int*>(DimsData());
+    for (int i = 0; i < size_; i++) {
+      buffer_size *= dims_data[i];
+    }
+    return buffer_size;
+  }
+
+ private:
+  // For use only by ExtendedShape(), written to guarantee (return-value) copy
+  // elision in C++17.
+  // This creates a shape padded to the desired size with the specified value.
+  RuntimeShape(int new_shape_size, const RuntimeShape& shape, int pad_value)
+      : size_(new_shape_size) {
+    // If the following check fails, it is likely because a 4D-only kernel is
+    // being used with an array of larger dimension count.
+    TFLITE_CHECK_GE(new_shape_size, shape.DimensionsCount());
+    const int size_increase = new_shape_size - shape.DimensionsCount();
+    for (int i = 0; i < size_increase; ++i) {
+      SetDim(i, pad_value);
+    }
+    std::memcpy(DimsData() + size_increase, shape.DimsData(),
+                sizeof(int32_t) * shape.DimensionsCount());
+  }
+
+  // A maximum of 4 dimensions are supported on TFLM.
+  static constexpr int kMaxSize = 5;
+  int32_t size_;
+  union {
+    int32_t dims_[kMaxSize];
+  };
+};
+
+// Since tensors with '0' in their shape are valid in TF, these offset functions
+// allow that as long as the corresponding index is also 0. It is upto the
+// calling ops to ensure that they perform verification checks on tensor shapes
+// if they don't support a particular behavior.
+
+inline int Offset(const RuntimeShape& shape, int i0, int i1, int i2, int i3) {
+  TFLITE_DCHECK_EQ(shape.DimensionsCount(), 4);
+  const int* dims_data = reinterpret_cast<const int*>(shape.DimsData());
+  TFLITE_DCHECK((dims_data[0] == 0 && i0 == 0) ||
+                (i0 >= 0 && i0 < dims_data[0]));
+  TFLITE_DCHECK((dims_data[1] == 0 && i1 == 0) ||
+                (i1 >= 0 && i1 < dims_data[1]));
+  TFLITE_DCHECK((dims_data[2] == 0 && i2 == 0) ||
+                (i2 >= 0 && i2 < dims_data[2]));
+  TFLITE_DCHECK((dims_data[3] == 0 && i3 == 0) ||
+                (i3 >= 0 && i3 < dims_data[3]));
+  return ((i0 * dims_data[1] + i1) * dims_data[2] + i2) * dims_data[3] + i3;
+}
+
+inline int Offset(const RuntimeShape& shape, int i0, int i1, int i2, int i3,
+                  int i4) {
+  TFLITE_DCHECK_EQ(shape.DimensionsCount(), 5);
+  const int* dims_data = reinterpret_cast<const int*>(shape.DimsData());
+  TFLITE_DCHECK((dims_data[0] == 0 && i0 == 0) ||
+                (i0 >= 0 && i0 < dims_data[0]));
+  TFLITE_DCHECK((dims_data[1] == 0 && i1 == 0) ||
+                (i1 >= 0 && i1 < dims_data[1]));
+  TFLITE_DCHECK((dims_data[2] == 0 && i2 == 0) ||
+                (i2 >= 0 && i2 < dims_data[2]));
+  TFLITE_DCHECK((dims_data[3] == 0 && i3 == 0) ||
+                (i3 >= 0 && i3 < dims_data[3]));
+  TFLITE_DCHECK((dims_data[4] == 0 && i4 == 0) ||
+                (i4 >= 0 && i4 < dims_data[4]));
+  return (((i0 * dims_data[1] + i1) * dims_data[2] + i2) * dims_data[3] + i3) *
+             dims_data[4] +
+         i4;
+}
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_RUNTIME_SHAPE_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/strided_slice_logic.h

@@ -0,0 +1,211 @@
+/* Copyright 2018 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_KERNELS_INTERNAL_STRIDED_SLICE_LOGIC_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_STRIDED_SLICE_LOGIC_H_
+
+#include <limits>
+#include <vector>
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace strided_slice {
+
+// Use until std::clamp() is available from C++17.
+inline int Clamp(const int v, const int lo, const int hi) {
+  TFLITE_DCHECK(!(hi < lo));
+  if (hi < v) return hi;
+  if (v < lo) return lo;
+  return v;
+}
+
+inline void StridedSlicePadIndices(tflite::StridedSliceParams* p,
+                                   int dim_count) {
+  // Add indices and mask bits to fully include extra dimensions
+  TFLITE_CHECK_LE(dim_count, 5);
+  TFLITE_CHECK_GE(dim_count, p->start_indices_count);
+  TFLITE_CHECK_EQ(p->start_indices_count, p->stop_indices_count);
+  TFLITE_CHECK_EQ(p->stop_indices_count, p->strides_count);
+
+  const int pad_count = dim_count - p->start_indices_count;
+
+  // Pad indices at start, so move arrays by pad_count.
+  for (int i = p->start_indices_count - 1; i >= 0; --i) {
+    p->strides[i + pad_count] = p->strides[i];
+    p->start_indices[i + pad_count] = p->start_indices[i];
+    p->stop_indices[i + pad_count] = p->stop_indices[i];
+  }
+  for (int i = 0; i < pad_count; ++i) {
+    p->start_indices[i] = 0;
+    p->stop_indices[i] = 1;
+    p->strides[i] = 1;
+  }
+
+  // Pad masks with 0s or 1s as required.
+  p->shrink_axis_mask <<= pad_count;
+  p->ellipsis_mask <<= pad_count;
+  p->new_axis_mask <<= pad_count;
+  p->begin_mask <<= pad_count;
+  p->end_mask <<= pad_count;
+  p->begin_mask |= (1 << pad_count) - 1;
+  p->end_mask |= (1 << pad_count) - 1;
+
+  p->start_indices_count = dim_count;
+  p->stop_indices_count = dim_count;
+  p->strides_count = dim_count;
+}
+
+// Return the index for the first element along that axis. This index will be a
+// positive integer between [0, axis_size] (or [-1, axis_size -1] if stride < 0)
+// that can be used to index directly into the data.
+inline int StartForAxis(const tflite::StridedSliceParams& params,
+                        const RuntimeShape& input_shape, int axis) {
+  const auto begin_mask = params.begin_mask;
+  const auto* start_indices = params.start_indices;
+  const auto* strides = params.strides;
+  const int axis_size = input_shape.Dims(axis);
+  if (axis_size == 0) {
+    return 0;
+  }
+  // Begin with the specified index.
+  int start = start_indices[axis];
+
+  // begin_mask override
+  if (begin_mask & 1 << axis) {
+    if (strides[axis] > 0) {
+      // Forward iteration - use the first element. These values will get
+      // clamped below (Note: We could have set them to 0 and axis_size-1, but
+      // use lowest() and max() to maintain symmetry with StopForAxis())
+      start = std::numeric_limits<int>::lowest();
+    } else {
+      // Backward iteration - use the last element.
+      start = std::numeric_limits<int>::max();
+    }
+  }
+
+  // Handle negative indices
+  if (start < 0) {
+    start += axis_size;
+  }
+
+  // Clamping
+  if (strides[axis] > 0) {
+    // Forward iteration
+    start = Clamp(start, 0, axis_size);
+  } else {
+    // Backward iteration
+    start = Clamp(start, -1, axis_size - 1);
+  }
+
+  return start;
+}
+
+// Return the "real" index for the end of iteration along that axis. This is an
+// "end" in the traditional C sense, in that it points to one past the last
+// element. ie. So if you were iterating through all elements of a 1D array of
+// size 4, this function would return 4 as the stop, because it is one past the
+// "real" indices of 0, 1, 2 & 3.
+inline int StopForAxis(const tflite::StridedSliceParams& params,
+                       const RuntimeShape& input_shape, int axis,
+                       int start_for_axis) {
+  const auto end_mask = params.end_mask;
+  const auto shrink_axis_mask = params.shrink_axis_mask;
+  const auto* stop_indices = params.stop_indices;
+  const auto* strides = params.strides;
+  const int axis_size = input_shape.Dims(axis);
+  if (axis_size == 0) {
+    return 0;
+  }
+
+  // Begin with the specified index
+  const bool shrink_axis = shrink_axis_mask & (1 << axis);
+  int stop = stop_indices[axis];
+
+  // When shrinking an axis, the end position does not matter (and can be
+  // incorrect when negative indexing is used, see Issue #19260). Always use
+  // start_for_axis + 1 to generate a length 1 slice, since start_for_axis has
+  // already been adjusted for negative indices.
+  if (shrink_axis) {
+    return start_for_axis + 1;
+  }
+
+  // end_mask override
+  if (end_mask & (1 << axis)) {
+    if (strides[axis] > 0) {
+      // Forward iteration - use the last element. These values will get
+      // clamped below
+      stop = std::numeric_limits<int>::max();
+    } else {
+      // Backward iteration - use the first element.
+      stop = std::numeric_limits<int>::lowest();
+    }
+  }
+
+  // Handle negative indices
+  if (stop < 0) {
+    stop += axis_size;
+  }
+
+  // Clamping
+  // Because the end index points one past the last element, we need slightly
+  // different clamping ranges depending on the direction.
+  if (strides[axis] > 0) {
+    // Forward iteration
+    stop = Clamp(stop, 0, axis_size);
+  } else {
+    // Backward iteration
+    stop = Clamp(stop, -1, axis_size - 1);
+  }
+
+  return stop;
+}
+
+inline bool LoopCondition(int index, int stop, int stride) {
+  // True when we have reached the end of an axis and should loop.
+  return stride > 0 ? index >= stop : index <= stop;
+}
+
+inline tflite::StridedSliceParams BuildStridedSliceParams(
+    int begin_mask, int end_mask, int shrink_axis_mask,
+    const std::vector<int>& start_indices, const std::vector<int>& stop_indices,
+    const std::vector<int>& strides) {
+  tflite::StridedSliceParams op_params;
+  const int dims_count = start_indices.size();
+
+  op_params.start_indices_count = dims_count;
+  op_params.stop_indices_count = dims_count;
+  op_params.strides_count = dims_count;
+  for (int i = 0; i < dims_count; ++i) {
+    op_params.start_indices[i] = start_indices[i];
+    op_params.stop_indices[i] = stop_indices[i];
+    op_params.strides[i] = strides[i];
+  }
+
+  op_params.begin_mask = begin_mask;
+  op_params.ellipsis_mask = 0;
+  op_params.end_mask = end_mask;
+  op_params.new_axis_mask = 0;
+  op_params.shrink_axis_mask = shrink_axis_mask;
+
+  return op_params;
+}
+
+}  // namespace strided_slice
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_STRIDED_SLICE_LOGIC_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/tensor_ctypes.h

@@ -0,0 +1,47 @@
+/* Copyright 2017 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_KERNELS_INTERNAL_TENSOR_CTYPES_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+template <typename T>
+inline T* GetTensorData(TfLiteTensor* tensor) {
+  return tensor != nullptr ? reinterpret_cast<T*>(tensor->data.raw) : nullptr;
+}
+
+template <typename T>
+inline const T* GetTensorData(const TfLiteTensor* tensor) {
+  return tensor != nullptr ? reinterpret_cast<const T*>(tensor->data.raw)
+                           : nullptr;
+}
+
+inline RuntimeShape GetTensorShape(const TfLiteTensor* tensor) {
+  if (tensor == nullptr) {
+    return RuntimeShape();
+  }
+
+  TfLiteIntArray* dims = tensor->dims;
+  const int dims_size = dims->size;
+  const int32_t* dims_data = reinterpret_cast<const int32_t*>(dims->data);
+  return RuntimeShape(dims_size, dims_data);
+}
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_TENSOR_CTYPES_H_