chore: import upstream snapshot with attribution

tensorflow/lite/kernels/internal/reference/add.h

@@ -0,0 +1,427 @@
+/* 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 <algorithm>
+#include <cstddef>
+#include <cstdint>
+#include <type_traits>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.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<T>(
+        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 void AddBroadcast(const T* input_data, const T* broadcast_data,
+                         T* output_data, size_t size, T activation_min,
+                         T activation_max) {
+  for (size_t c = 0; c < size; ++c) {
+    output_data[c] = ActivationFunctionWithMinMax<T>(
+        input_data[c] + broadcast_data[0], activation_min, activation_max);
+  }
+}
+
+template <>
+inline void AddBroadcast<int32_t>(const int32_t* input_data,
+                                  const int32_t* broadcast_data,
+                                  int32_t* output_data, size_t size,
+                                  int32_t activation_min,
+                                  int32_t activation_max) {
+  size_t c = 0;
+#ifdef USE_NEON
+  const int32x4_t vmax = vdupq_n_s32(activation_max);
+  const int32x4_t vmin = vdupq_n_s32(activation_min);
+  const int32x4_t vb = vdupq_n_s32(broadcast_data[0]);
+  for (; c + 4 <= size; c += 4) {
+    const int32x4_t va = vld1q_s32(&input_data[c]);
+    int32x4_t vres = vaddq_s32(va, vb);
+    vres = vmaxq_s32(vmin, vres);
+    vres = vminq_s32(vmax, vres);
+    vst1q_s32(&output_data[c], vres);
+  }
+#endif
+  for (; c < size; ++c) {
+    output_data[c] = ActivationFunctionWithMinMax<int32_t>(
+        input_data[c] + broadcast_data[0], activation_min, activation_max);
+  }
+}
+
+template <typename T>
+void AddElementwise(const T* input1_data, const T* input2_data, T* output_data,
+                    size_t size, T activation_min, T activation_max) {
+  for (size_t c = 0; c < size; ++c) {
+    output_data[c] = ActivationFunctionWithMinMax<T>(
+        input1_data[c] + input2_data[c], activation_min, activation_max);
+  }
+}
+
+template <>
+inline void AddElementwise<int32_t>(const int32_t* input1_data,
+                                    const int32_t* input2_data,
+                                    int32_t* output_data, size_t size,
+                                    int32_t activation_min,
+                                    int32_t activation_max) {
+  size_t c = 0;
+#ifdef USE_NEON
+  const int32x4_t vmax = vdupq_n_s32(activation_max);
+  const int32x4_t vmin = vdupq_n_s32(activation_min);
+  for (; c + 4 <= size; c += 4) {
+    const int32x4_t va = vld1q_s32(&input1_data[c]);
+    const int32x4_t vb = vld1q_s32(&input2_data[c]);
+    int32x4_t vres = vaddq_s32(va, vb);
+    vres = vmaxq_s32(vmin, vres);
+    vres = vminq_s32(vmax, vres);
+    vst1q_s32(&output_data[c], vres);
+  }
+#endif
+  for (; c < size; ++c) {
+    output_data[c] = ActivationFunctionWithMinMax<int32_t>(
+        input1_data[c] + input2_data[c], activation_min, activation_max);
+  }
+}
+
+template <typename T,
+          // For unquantized add for small integers, explicitly set to true.
+          bool dummy = false>
+inline typename std::enable_if<!is_small_integer<T>::value || dummy, void>::type
+BroadcastAdd6DSlow(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);
+  auto op = [activation_min, activation_max](T a, T b) {
+    return ActivationFunctionWithMinMax<T>(a + b, activation_min,
+                                           activation_max);
+  };
+  BroadcastBinaryOpSimple(input1_shape, input1_data, input2_shape, input2_data,
+                          output_shape, output_data, op);
+}
+
+// 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
+BroadcastAdd6DSlow(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) {
+  auto op = [&params](T a, T b) {
+    const int32_t input1_val = params.input1_offset + a;
+    const int32_t input2_val = params.input2_offset + b;
+    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<T>(clamped_output);
+  };
+  BroadcastBinaryOpSimple(input1_shape, input1_data, input2_shape, input2_data,
+                          output_shape, output_data, op);
+}
+
+template <typename T>
+inline void 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) {
+  return BroadcastAdd6DSlow(params, input1_shape, input1_data, input2_shape,
+                            input2_data, output_shape, output_data);
+}
+
+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_

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_

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_

tensorflow/lite/kernels/internal/reference/batch_matmul.h

@@ -0,0 +1,289 @@
+/* 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 <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/portable_tensor_utils.h"
+#include "tensorflow/lite/kernels/internal/runtime_shape.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 size_t extent(const RuntimeShape& shape, int x) {
+  if (shape.Dims(x) == 1) {
+    return 0;
+  }
+  size_t prod = 1;
+  for (int i = x + 1; i < shape.DimensionsCount(); ++i) {
+    prod *= static_cast<size_t>(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 size_t batch_dim0 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0)));
+  const size_t batch_dim1 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1)));
+  const size_t batch_dim2 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2)));
+
+  const size_t lhs_ext0 = batch_matmul::extent(extended_lhs_shape, 0);
+  const size_t lhs_ext1 = batch_matmul::extent(extended_lhs_shape, 1);
+  const size_t lhs_ext2 = batch_matmul::extent(extended_lhs_shape, 2);
+  const size_t rhs_ext0 = batch_matmul::extent(extended_rhs_shape, 0);
+  const size_t rhs_ext1 = batch_matmul::extent(extended_rhs_shape, 1);
+  const size_t rhs_ext2 = batch_matmul::extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const size_t lhs_rows = static_cast<size_t>(extended_lhs_shape.Dims(3));
+  const size_t rhs_cols = static_cast<size_t>(extended_rhs_shape.Dims(4));
+  const size_t accum_depth = static_cast<size_t>(extended_lhs_shape.Dims(4));
+
+  for (size_t 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 (size_t 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 (size_t 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 (size_t j = 0; j < rhs_cols; ++j) {
+          for (size_t i = 0; i < lhs_rows; ++i) {
+            Tout total = 0;
+            for (size_t 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]);
+            }
+            size_t 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 float* per_channel_scales) {
+  const RuntimeShape extended_lhs_shape =
+      RuntimeShape::ExtendedShape(5, lhs_shape);
+  const RuntimeShape extended_rhs_shape =
+      RuntimeShape::ExtendedShape(5, rhs_shape);
+
+  const size_t batch_dim0 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0)));
+  const size_t batch_dim1 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1)));
+  const size_t batch_dim2 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2)));
+
+  const size_t lhs_ext0 = batch_matmul::extent(extended_lhs_shape, 0);
+  const size_t lhs_ext1 = batch_matmul::extent(extended_lhs_shape, 1);
+  const size_t lhs_ext2 = batch_matmul::extent(extended_lhs_shape, 2);
+  const size_t rhs_ext0 = batch_matmul::extent(extended_rhs_shape, 0);
+  const size_t rhs_ext1 = batch_matmul::extent(extended_rhs_shape, 1);
+  const size_t rhs_ext2 = batch_matmul::extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const size_t lhs_rows = static_cast<size_t>(extended_lhs_shape.Dims(3));
+  const size_t rhs_cols = static_cast<size_t>(extended_rhs_shape.Dims(4));
+  const size_t accum_depth = static_cast<size_t>(extended_lhs_shape.Dims(4));
+
+  const size_t ioff_ext0 = rhs_ext0 == 0 ? 0 : rhs_cols;
+  const size_t ioff_ext1 = rhs_ext1 == 0 ? 0 : rhs_cols;
+  const size_t ioff_ext2 = rhs_ext2 == 0 ? 0 : rhs_cols;
+  const size_t woff_ext0 = lhs_ext0 == 0 ? 0 : lhs_rows;
+  const size_t woff_ext1 = lhs_ext1 == 0 ? 0 : lhs_rows;
+  const size_t woff_ext2 = lhs_ext2 == 0 ? 0 : lhs_rows;
+
+  if (!compute_row_sums || *compute_row_sums) {
+    size_t num_weights_matrices = 1;
+    for (int i = 1; i < extended_lhs_shape.DimensionsCount() - 2; ++i) {
+      num_weights_matrices *= static_cast<size_t>(extended_lhs_shape.Dims(i));
+    }
+    TFLITE_DCHECK_LE(num_weights_matrices * lhs_rows,
+                     static_cast<size_t>(std::numeric_limits<int>::max()));
+    TFLITE_DCHECK_LE(accum_depth,
+                     static_cast<size_t>(std::numeric_limits<int>::max()));
+    tensor_utils::ReductionSumVector(
+        lhs_data, row_sums, static_cast<int>(num_weights_matrices * lhs_rows),
+        static_cast<int>(accum_depth));
+    if (compute_row_sums) {
+      *compute_row_sums = false;
+    }
+  }
+
+  for (size_t 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 (size_t 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 (size_t 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 (size_t 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 (size_t i = 0; i < lhs_rows; ++i) {
+            int32_t total = 0;
+            for (size_t 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;
+            size_t idx = lhs_rows * j + i;
+            float scale = batch_scaling_factor;
+            if (per_channel_scales) {
+              scale *= per_channel_scales[i];
+            }
+            out_ptr[idx] += scale * total;
+          }
+        }
+      }
+    }
+  }
+}
+
+template <typename lhsT, typename AccumT, typename rhsT = lhsT,
+          typename outputT = lhsT>
+inline void BatchMatMul(const FullyConnectedParams& params,
+                        const RuntimeShape& lhs_shape, const lhsT* lhs_data,
+                        const RuntimeShape& rhs_shape, const rhsT* rhs_data,
+                        const RuntimeShape& output_shape,
+                        outputT* output_data) {
+  const RuntimeShape extended_lhs_shape =
+      RuntimeShape::ExtendedShape(5, lhs_shape);
+  const RuntimeShape extended_rhs_shape =
+      RuntimeShape::ExtendedShape(5, rhs_shape);
+
+  const size_t batch_dim0 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(0), extended_rhs_shape.Dims(0)));
+  const size_t batch_dim1 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(1), extended_rhs_shape.Dims(1)));
+  const size_t batch_dim2 = static_cast<size_t>(batch_matmul::broadcast_dim(
+      extended_lhs_shape.Dims(2), extended_rhs_shape.Dims(2)));
+
+  const size_t lhs_ext0 = batch_matmul::extent(extended_lhs_shape, 0);
+  const size_t lhs_ext1 = batch_matmul::extent(extended_lhs_shape, 1);
+  const size_t lhs_ext2 = batch_matmul::extent(extended_lhs_shape, 2);
+  const size_t rhs_ext0 = batch_matmul::extent(extended_rhs_shape, 0);
+  const size_t rhs_ext1 = batch_matmul::extent(extended_rhs_shape, 1);
+  const size_t rhs_ext2 = batch_matmul::extent(extended_rhs_shape, 2);
+
+  // Set params for each matrix multiply.
+  const size_t lhs_rows = static_cast<size_t>(extended_lhs_shape.Dims(3));
+  const size_t rhs_cols = static_cast<size_t>(extended_rhs_shape.Dims(4));
+  const size_t accum_depth = static_cast<size_t>(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 (size_t b0 = 0; b0 < batch_dim0; ++b0) {
+    const lhsT* lhs_ptr0 = lhs_data + (b0 * lhs_ext0);
+    const rhsT* rhs_ptr0 = rhs_data + (b0 * rhs_ext0);
+    for (size_t b1 = 0; b1 < batch_dim1; ++b1) {
+      const lhsT* lhs_ptr1 = lhs_ptr0 + b1 * lhs_ext1;
+      const rhsT* rhs_ptr1 = rhs_ptr0 + b1 * rhs_ext1;
+      for (size_t b2 = 0; b2 < batch_dim2; ++b2) {
+        const lhsT* lhs_ptr2 = lhs_ptr1 + b2 * lhs_ext2;
+        const rhsT* rhs_ptr2 = rhs_ptr1 + b2 * rhs_ext2;
+        outputT* out_ptr = output_data + ((b0 * batch_dim1 * batch_dim2) +
+                                          b1 * batch_dim2 + b2) *
+                                             lhs_rows * rhs_cols;
+
+        for (size_t j = 0; j < rhs_cols; ++j) {
+          for (size_t i = 0; i < lhs_rows; ++i) {
+            AccumT total = 0;
+            for (size_t 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 size_t idx = lhs_rows * j + i;
+            out_ptr[idx] = static_cast<outputT>(total_scaled);
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BATCH_MATMUL_H_

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_

tensorflow/lite/kernels/internal/reference/binary_function.h

@@ -0,0 +1,60 @@
+/* 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/reference/broadcast_loop.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)) {
+  auto op = [func](T1 a, T2 b) { return func(a, b); };
+  BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                          unextended_input2_shape, input2_data,
+                          unextended_output_shape, output_data, op);
+}
+
+// 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_

tensorflow/lite/kernels/internal/reference/broadcast_args.h

@@ -0,0 +1,56 @@
+/* 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_BROADCAST_ARGS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BROADCAST_ARGS_H_
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+void BroadcastArgs(const RuntimeShape& input1_shape, const T* input1_data,
+                   const RuntimeShape& input2_shape, const T* input2_data,
+                   const RuntimeShape& output_shape, T* output_data) {
+  // Gets data at the backward index i of the shape tensor. Returns 1 if the
+  // index is out of range.
+  auto get_shape_data = [](const RuntimeShape& shape, const T* data,
+                           int backward_idx) -> T {
+    int forward_idx = shape.FlatSize() - 1 - backward_idx;
+    if (forward_idx < 0) return 1;
+    return data[forward_idx];
+  };
+
+  int output_num_elements = output_shape.FlatSize();
+  for (int i = 0; i < output_num_elements; ++i) {
+    int backward_i = output_num_elements - 1 - i;
+    int shape1_i = get_shape_data(input1_shape, input1_data, i);
+    int shape2_i = get_shape_data(input2_shape, input2_data, i);
+    if (shape1_i == 1) {
+      output_data[backward_i] = shape2_i;
+    } else if (shape2_i == 1) {
+      output_data[backward_i] = shape1_i;
+    } else {
+      TFLITE_CHECK_EQ(shape1_i, shape2_i);
+      output_data[backward_i] = shape1_i;
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BROADCAST_ARGS_H_

tensorflow/lite/kernels/internal/reference/broadcast_loop.h

@@ -0,0 +1,172 @@
+/* Copyright 2026 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_BROADCAST_LOOP_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BROADCAST_LOOP_H_
+
+#include <algorithm>
+#include <cstddef>
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/runtime_shape.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename Pointer1, typename Pointer2, typename OutputType,
+          typename BinaryOp>
+void RunBinaryOp(Pointer1 a, Pointer2 b, OutputType* output,
+                 const size_t* a_stride, const size_t* b_stride,
+                 const size_t* output_stride, const size_t* output_shape,
+                 int dim, BinaryOp op) {
+  TFLITE_DCHECK_GE(dim, 0);
+  size_t output_shape_0 = output_shape[dim];
+  size_t output_stride_0 = output_stride[dim];
+  size_t a_stride_0 = a_stride[dim];
+  size_t b_stride_0 = b_stride[dim];
+  if (dim == 0) {
+    TFLITE_DCHECK_EQ(output_stride_0, 1);
+    if (a_stride_0 == 0) {
+      TFLITE_DCHECK_EQ(b_stride_0, 1);
+      const auto a_0 = *a;
+      for (size_t i = 0; i < output_shape_0; ++i) {
+        output[i] = op(a_0, b[i]);
+      }
+    } else if (b_stride_0 == 0) {
+      TFLITE_DCHECK_EQ(a_stride_0, 1);
+      const auto b_0 = *b;
+      for (size_t i = 0; i < output_shape_0; ++i) {
+        output[i] = op(a[i], b_0);
+      }
+    } else {
+      TFLITE_DCHECK_EQ(a_stride_0, 1);
+      TFLITE_DCHECK_EQ(b_stride_0, 1);
+      for (size_t i = 0; i < output_shape_0; ++i) {
+        output[i] = op(a[i], b[i]);
+      }
+    }
+  } else {
+    dim -= 1;
+    for (size_t i = 0; i < output_shape_0; ++i) {
+      RunBinaryOp(a, b, output, a_stride, b_stride, output_stride, output_shape,
+                  dim, op);
+      a = a + a_stride_0;
+      b = b + b_stride_0;
+      output = output + output_stride_0;
+    }
+  }
+}
+
+// Returns true if a dimension of a loop nest can be fused with the previous
+// dimension in the loop nest.
+inline bool CanFuseLoops(size_t output_dim, size_t dim, size_t stride,
+                         size_t expected_stride, size_t next_stride) {
+  if (output_dim == 1) {
+    // We can always fuse an extent 1 dimension.
+    return true;
+  }
+  if (next_stride == 0) {
+    // The next loop's stride is 0, the current stride must be 0 too.
+    return stride == 0;
+  } else {
+    // The next loop's stride must match the current loop's stride.
+    return stride == expected_stride;
+  }
+}
+
+template <typename Pointer1, typename Pointer2, typename OutputType,
+          typename BinaryOp>
+inline void BroadcastBinaryOpSimple(const RuntimeShape& input1_shape,
+                                    Pointer1 input1_data,
+                                    const RuntimeShape& input2_shape,
+                                    Pointer2 input2_data,
+                                    const RuntimeShape& output_shape,
+                                    OutputType* output_data, BinaryOp op) {
+  constexpr int kMaxRank = 8;
+
+  int dims_count = std::max(
+      output_shape.DimensionsCount(),
+      std::max(input1_shape.DimensionsCount(), input2_shape.DimensionsCount()));
+
+  if (dims_count <= 0) {
+    *output_data = op(*input1_data, *input2_data);
+    return;
+  }
+
+  TFLITE_DCHECK_LE(dims_count, kMaxRank);
+
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(dims_count, output_shape);
+  const RuntimeShape extended_input1_shape =
+      RuntimeShape::ExtendedShape(dims_count, input1_shape);
+  const RuntimeShape extended_input2_shape =
+      RuntimeShape::ExtendedShape(dims_count, input2_shape);
+
+  // This loop works as follows:
+  // - Start with the last loop, the stride 1 dimension, populate the innermost
+  //   loop (dimension 0).
+  // - For all remaining dimensions, if the loop can be fused with the previous
+  //   loop do that.
+  // - Otherwise, make a new loop.
+  size_t a_strides[kMaxRank];
+  size_t b_strides[kMaxRank];
+  size_t o_strides[kMaxRank];
+  size_t o_shape[kMaxRank];
+
+  size_t a_accum_stride = 1;
+  size_t b_accum_stride = 1;
+  size_t o_accum_stride = 1;
+  int next_dim_idx = -1;
+  for (int i = dims_count - 1; i >= 0; --i) {
+    const int input1_dim = extended_input1_shape.Dims(i);
+    const int input2_dim = extended_input2_shape.Dims(i);
+    const int output_dim = extended_output_shape.Dims(i);
+    if (input1_dim <= 0 || input2_dim <= 0 || output_dim <= 0) {
+      // Empty operation.
+      return;
+    }
+    size_t a_stride = (input1_dim == 1 && output_dim != 1) ? 0 : a_accum_stride;
+    size_t b_stride = (input2_dim == 1 && output_dim != 1) ? 0 : b_accum_stride;
+    size_t o_stride = o_accum_stride;
+    if (next_dim_idx >= 0 &&
+        CanFuseLoops(output_dim, input1_dim, a_stride, a_accum_stride,
+                     a_strides[next_dim_idx]) &&
+        CanFuseLoops(output_dim, input2_dim, b_stride, b_accum_stride,
+                     b_strides[next_dim_idx]) &&
+        CanFuseLoops(output_dim, output_dim, o_stride, o_accum_stride,
+                     o_strides[next_dim_idx])) {
+      // This dimension can be fused into one loop with the previous
+      // dimension.
+      o_shape[next_dim_idx] *= output_dim;
+    } else {
+      ++next_dim_idx;
+      a_strides[next_dim_idx] = a_stride;
+      b_strides[next_dim_idx] = b_stride;
+      o_strides[next_dim_idx] = o_stride;
+      o_shape[next_dim_idx] = output_dim;
+    }
+
+    a_accum_stride *= input1_dim;
+    b_accum_stride *= input2_dim;
+    o_accum_stride *= output_dim;
+  }
+
+  RunBinaryOp(input1_data, input2_data, output_data, a_strides, b_strides,
+              o_strides, o_shape, next_dim_idx, op);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BROADCAST_LOOP_H_

tensorflow/lite/kernels/internal/reference/broadcast_to.h

@@ -0,0 +1,147 @@
+/* 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_BROADCAST_TO_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BROADCAST_TO_H_
+
+#include <cstddef>
+#include <vector>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/kernel_util.h"
+
+namespace tflite {
+namespace reference_ops {
+template <int N>
+void BroadcastImpl(const NdArrayDesc<N>& input_desc, const char* input_data,
+                   const NdArrayDesc<N>& output_desc, char* output_data,
+                   int indexes[N], int dim, const int last_broadcasting_dim,
+                   const int type_size) {
+  // Copy data from input to output.
+  if (dim == last_broadcasting_dim) {
+    size_t copy_size =
+        static_cast<size_t>(output_desc.strides[dim]) * type_size;
+    const char* data_src =
+        input_data +
+        static_cast<size_t>(SubscriptToIndex(input_desc, indexes)) * type_size;
+    char* data_dst =
+        output_data +
+        static_cast<size_t>(SubscriptToIndex(output_desc, indexes)) * type_size;
+    for (int i = 0; i < output_desc.extents[dim]; ++i, data_dst += copy_size) {
+      memcpy(data_dst, data_src, copy_size);
+    }
+    return;
+  }
+
+  // Recursive call to find the next broadcasting.
+  for (indexes[dim] = 0; indexes[dim] < input_desc.extents[dim];
+       ++indexes[dim]) {
+    BroadcastImpl<N>(input_desc, input_data, output_desc, output_data, indexes,
+                     dim + 1, last_broadcasting_dim, type_size);
+  }
+
+  // Duplicate data in output tensor.
+  indexes[dim] = 0;
+  if (input_desc.extents[dim] != output_desc.extents[dim]) {
+    size_t copy_size =
+        static_cast<size_t>(output_desc.strides[dim] * type_size);
+    char* data_src =
+        output_data +
+        static_cast<size_t>(SubscriptToIndex(output_desc, indexes)) * type_size;
+    char* data_dst = data_src + copy_size;
+    for (int i = 1; i < output_desc.extents[dim]; ++i, data_dst += copy_size) {
+      memcpy(data_dst, data_src, copy_size);
+    }
+  }
+}
+
+template <int N>
+inline void BroadcastTo(const RuntimeShape& unextended_input_shape,
+                        const char* input_data,
+                        const RuntimeShape& unextended_output_shape,
+                        char* output_data, TfLiteType data_type) {
+  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);
+
+  // Get the last dimension has broadcasting. At this dimension, the data is
+  // copied from input tensor to output tensor.
+  int last_broadcast_dim = -1;
+  for (int i = N - 1; i >= 0; --i) {
+    if (input_desc.extents[i] != output_desc.extents[i]) {
+      last_broadcast_dim = i;
+      break;
+    }
+  }
+
+  // If non-broadcasting, just copy data from input to output tensor.
+  if (last_broadcast_dim == -1) {
+    memcpy(output_data, input_data,
+           static_cast<size_t>(unextended_input_shape.FlatSize()) *
+               static_cast<size_t>(TfLiteTypeGetSize(data_type)));
+    return;
+  }
+
+  // Broadcasting using memcpy.
+  int indexes[N] = {0};
+  BroadcastImpl<N>(input_desc, input_data, output_desc, output_data, indexes, 0,
+                   last_broadcast_dim, TfLiteTypeGetSize(data_type));
+}
+
+inline void BroadcastTo(const RuntimeShape& unextended_input_shape,
+                        const char* input_data,
+                        const RuntimeShape& unextended_output_shape,
+                        char* output_data, TfLiteType data_type) {
+  const int dims = unextended_output_shape.DimensionsCount();
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(dims, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(dims, unextended_output_shape);
+  const int type_size = TfLiteTypeGetSize(data_type);
+  if (dims == 0) {
+    memcpy(output_data, input_data, type_size);
+    return;
+  }
+
+  std::vector<int> input_strides(dims);
+  std::vector<int> output_strides(dims);
+  input_strides[dims - 1] = 1;
+  output_strides[dims - 1] = 1;
+  for (int i = dims - 2; i >= 0; --i) {
+    input_strides[i] = input_strides[i + 1] * input_shape.Dims(i + 1);
+    output_strides[i] = output_strides[i + 1] * output_shape.Dims(i + 1);
+  }
+
+  const int output_flat_size = unextended_output_shape.FlatSize();
+  for (int output_index = 0; output_index < output_flat_size; ++output_index) {
+    int remaining_index = output_index;
+    int input_index = 0;
+    for (int dim = 0; dim < dims; ++dim) {
+      const int coordinate = remaining_index / output_strides[dim];
+      remaining_index %= output_strides[dim];
+      if (input_shape.Dims(dim) != 1) {
+        input_index += coordinate * input_strides[dim];
+      }
+    }
+    memcpy(output_data + static_cast<size_t>(output_index) * type_size,
+           input_data + static_cast<size_t>(input_index) * type_size,
+           type_size);
+  }
+}
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BROADCAST_TO_H_

tensorflow/lite/kernels/internal/reference/cast.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_CAST_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CAST_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename SrcT, typename DstT>
+inline void Cast(const RuntimeShape& input_shape, const SrcT* input_data,
+                 const RuntimeShape& output_shape, DstT* 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] = static_cast<DstT>(input_data[offset]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CAST_H_

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_

tensorflow/lite/kernels/internal/reference/comparisons.h

@@ -0,0 +1,239 @@
+/* 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 <cstdint>
+
+#include "tensorflow/lite/core/c/common.h"
+#include "tensorflow/lite/core/macros.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.h"
+#include "tensorflow/lite/kernels/internal/runtime_shape.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);
+  }
+}
+
+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) {
+  BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                          unextended_input2_shape, input2_data,
+                          unextended_output_shape, output_data, F);
+}
+
+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) {
+  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;
+
+  auto op = [=](T a, T b) {
+    const int32_t input1_val = input1_offset + a;
+    const int32_t input2_val = input2_offset + b;
+    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);
+    return F(scaled_input1_val, scaled_input2_val);
+  };
+
+  BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                          unextended_input2_shape, input2_data,
+                          unextended_output_shape, output_data, op);
+}
+
+#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_

tensorflow/lite/kernels/internal/reference/concatenation.h

@@ -0,0 +1,218 @@
+/* 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 <algorithm>
+#include <cstddef>
+
+#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 (int64_t k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int64_t 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;
+    }
+  }
+}
+
+template <>
+inline void Concatenation<Int4>(const ConcatenationParams& params,
+                                const RuntimeShape* const* input_shapes,
+                                const Int4* const* input_data,
+                                const RuntimeShape& output_shape,
+                                Int4* 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);
+  }
+
+  uint8_t* output_ptr = reinterpret_cast<uint8_t*>(output_data);
+  // We can't guarantee that the output buffer is initialized to 0, so we have
+  // to clear it to ensure the high/low nibbles not currently being written are
+  // not garbage.
+  // Note: The total number of elements (nibbles) is outer_size *
+  // output_shape.Dims(axis) * base_inner_size. We use int64_t to avoid
+  // overflow issues with FlatSize().
+  int64_t total_elements =
+      outer_size * output_shape.Dims(axis) * base_inner_size;
+  // Bytes needed: (elements + 1) / 2.
+  memset(output_ptr, 0, (static_cast<size_t>(total_elements) + 1) / 2);
+
+  int64_t output_offset = 0;
+  for (int64_t k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int64_t copy_size = input_shapes[i]->Dims(axis) * base_inner_size;
+      const uint8_t* input_ptr =
+          reinterpret_cast<const uint8_t*>(input_data[i]);
+      // The input_ptr points to the start of the tensor data.
+      // We need to calculate the offset for the current outer loop iteration
+      // 'k'.
+      // The tensor has total elements = outer_size * copy_size.
+      // So current offset in elements is k * copy_size.
+      int64_t input_offset = k * copy_size;
+
+      for (int64_t j = 0; j < copy_size; ++j) {
+        int64_t in_idx = input_offset + j;
+        uint8_t val = input_ptr[in_idx / 2];
+        uint8_t nibble = (in_idx % 2 == 0) ? (val & 0x0F) : ((val >> 4) & 0x0F);
+
+        int64_t out_idx = output_offset + j;
+        uint8_t* out_byte = output_ptr + (out_idx / 2);
+        if (out_idx % 2 == 0) {
+          *out_byte = (*out_byte & 0xF0) | nibble;
+        } else {
+          *out_byte = (*out_byte & 0x0F) | (nibble << 4);
+        }
+      }
+      output_offset += 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 (int64_t k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int64_t 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 (int64_t 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_

tensorflow/lite/kernels/internal/reference/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_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_
+
+#include <algorithm>
+
+#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 = input_shape.Dims(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 filter_input_depth = filter_shape.Dims(3);
+  const int groups = input_depth / filter_input_depth;
+  TFLITE_DCHECK_NE(groups, 0);
+  TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
+  const int filters_per_group = output_depth / groups;
+  TFLITE_DCHECK_NE(filters_per_group, 0);
+  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) {
+          auto group = out_channel / filters_per_group;
+          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 < filter_input_depth;
+                   ++in_channel) {
+                float input_value =
+                    input_data[Offset(input_shape, batch, in_y, in_x,
+                                      in_channel + group * filter_input_depth)];
+                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 = input_shape.Dims(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 filter_input_depth = filter_shape.Dims(3);
+  const int groups = input_depth / filter_input_depth;
+  TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
+  const int filters_per_group = output_depth / groups;
+  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) {
+          auto group = out_channel / filters_per_group;
+          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 < filter_input_depth;
+                   ++in_channel) {
+                int32_t input_val =
+                    input_data[Offset(input_shape, batch, in_y, in_x,
+                                      in_channel + group * filter_input_depth)];
+                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 = input_shape.Dims(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 filter_input_depth = filter_shape.Dims(3);
+  const int groups = input_depth / filter_input_depth;
+  TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
+  const int filters_per_group = output_depth / groups;
+  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) {
+          auto group = out_channel / filters_per_group;
+          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 < filter_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 + group * filter_input_depth)];
+                  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_

tensorflow/lite/kernels/internal/reference/conv3d.h

@@ -0,0 +1,114 @@
+/* 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_CONV3D_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV3D_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void Conv3D(const Conv3DParams& 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) {
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 5);
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_num_channels = MatchingDim(input_shape, 4, filter_shape, 3);
+  const int output_num_channels = MatchingDim(filter_shape, 4, output_shape, 4);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_num_channels);
+  }
+
+  // Only NDHWC format is currently supported.
+  const int input_width = input_shape.Dims(3);
+  const int input_height = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_depth = filter_shape.Dims(0);
+  const int output_width = output_shape.Dims(3);
+  const int output_height = output_shape.Dims(2);
+  const int output_depth = output_shape.Dims(1);
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int pad_depth = params.padding_values.depth;
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_d = 0; out_d < output_depth; ++out_d) {
+      const int in_d_origin = (out_d * params.stride_depth) - pad_depth;
+      for (int out_y = 0; out_y < output_height; ++out_y) {
+        const int in_y_origin = (out_y * params.stride_height) - pad_height;
+        for (int out_x = 0; out_x < output_width; ++out_x) {
+          const int in_x_origin = (out_x * params.stride_width) - pad_width;
+          for (int out_channel = 0; out_channel < output_num_channels;
+               ++out_channel) {
+            float total = 0.f;
+            for (int filter_d = 0; filter_d < filter_depth; ++filter_d) {
+              const int in_d = in_d_origin + params.dilation_depth * filter_d;
+              for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+                const int in_y =
+                    in_y_origin + params.dilation_height * filter_y;
+                for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                  const int in_x =
+                      in_x_origin + params.dilation_width * 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) && (in_d >= 0) &&
+                      (in_d < input_depth);
+
+                  if (!is_point_inside_image) {
+                    continue;
+                  }
+
+                  for (int in_channel = 0; in_channel < input_num_channels;
+                       ++in_channel) {
+                    float input_value = input_data[Offset(
+                        input_shape, batch, in_d, in_y, in_x, in_channel)];
+                    float filter_value =
+                        filter_data[Offset(filter_shape, filter_d, filter_y,
+                                           filter_x, in_channel, out_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_d, out_y, out_x,
+                               out_channel)] =
+                ActivationFunctionWithMinMax(total + bias_value,
+                                             params.float_activation_min,
+                                             params.float_activation_max);
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV3D_H_

tensorflow/lite/kernels/internal/reference/conv3d_transpose.h

@@ -0,0 +1,134 @@
+/* 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_CONV3D_TRANSPOSE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV3D_TRANSPOSE_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void Conv3DTranspose(
+    const Conv3DTransposeParams& 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 stride_depth = params.stride_depth;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int pad_depth = params.padding_values.depth;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 5);
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_num_channels = MatchingDim(input_shape, 4, filter_shape, 4);
+  const int output_num_channels = output_shape.Dims(4);
+  const int input_depth = input_shape.Dims(1);
+  const int input_height = input_shape.Dims(2);
+  const int input_width = input_shape.Dims(3);
+  const int filter_depth = filter_shape.Dims(0);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_depth = output_shape.Dims(1);
+  const int output_height = output_shape.Dims(2);
+  const int output_width = output_shape.Dims(3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_num_channels);
+  }
+
+  // Initializes 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_d = 0; in_d < input_depth; ++in_d) {
+      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_num_channels;
+               ++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;
+            const int out_d_origin = (in_d * stride_depth) - pad_depth;
+            for (int filter_d = 0; filter_d < filter_depth; ++filter_d) {
+              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_num_channels;
+                       ++out_channel) {
+                    // Compute output element location.
+                    const int out_x =
+                        out_x_origin + params.dilation_width * filter_x;
+                    const int out_y =
+                        out_y_origin + params.dilation_height * filter_y;
+                    const int out_d =
+                        out_d_origin + params.dilation_depth * filter_d;
+                    // We cannot accumulate out of bounds.
+                    if ((out_x >= 0) && (out_x < output_width) &&
+                        (out_y >= 0) && (out_y < output_height) &&
+                        (out_d >= 0) && (out_d < output_depth)) {
+                      float input_value = input_data[Offset(
+                          input_shape, batch, in_d, in_y, in_x, in_channel)];
+                      float filter_value = filter_data[Offset(
+                          filter_shape, filter_d, filter_y, filter_x,
+                          out_channel, in_channel)];
+                      output_data[Offset(output_shape, batch, out_d, out_y,
+                                         out_x, out_channel)] +=
+                          input_value * filter_value;
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  const float float_activation_min = params.float_activation_min;
+  const float float_activation_max = params.float_activation_max;
+  float* data_ptr = output_data;
+  if (bias_data) {
+    const int outer_size =
+        batches * output_depth * output_height * output_width;
+    for (int n = 0; n < outer_size; ++n) {
+      for (int c = 0; c < output_num_channels; ++c) {
+        data_ptr[c] = ActivationFunctionWithMinMax(data_ptr[c] + bias_data[c],
+                                                   float_activation_min,
+                                                   float_activation_max);
+      }
+      data_ptr += output_num_channels;
+    }
+  } else {
+    const int flat_size = output_shape.FlatSize();
+    for (int i = 0; i < flat_size; ++i) {
+      data_ptr[i] = ActivationFunctionWithMinMax(
+          data_ptr[i], float_activation_min, float_activation_max);
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV3D_TRANSPOSE_H_

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_

tensorflow/lite/kernels/internal/reference/densify.h

@@ -0,0 +1,47 @@
+/* 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_DENSIFY_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DENSIFY_H_
+
+#include <vector>
+
+#include "tensorflow/lite/core/c/common.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/kernels/internal/utils/sparsity_format_converter.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T>
+inline void Densify(const TfLiteSparsity* sparsity,
+                    const RuntimeShape& input_shape, const T* input_data,
+                    const RuntimeShape& output_shape, T* output_data,
+                    TfLiteContext* context) {
+  const int dims_count = output_shape.DimensionsCount();
+  std::vector<int> vector_shape(dims_count);
+  for (int i = 0; i < dims_count; i++) {
+    vector_shape[i] = output_shape.Dims(i);
+  }
+
+  tflite::internal::sparsity::FormatConverter<T> converter(vector_shape,
+                                                           *sparsity);
+  converter.SparseToDense(input_data, output_shape.FlatSize(), output_data,
+                          context);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DENSIFY_H_

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_

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_

tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.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_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 int64_t oc =
+                  m + static_cast<int64_t>(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 int64_t output_channel =
+                  m + static_cast<int64_t>(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_

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_

tensorflow/lite/kernels/internal/reference/div.h

@@ -0,0 +1,235 @@
+/* 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_DIV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DIV_H_
+
+#include <algorithm>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void DivCheckArithmeticParams(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.
+  constexpr int32_t max_value =
+      (static_cast<int32_t>(std::numeric_limits<T>::max()) + 1);
+  TFLITE_DCHECK_GE(params.input1_offset, -max_value);
+  TFLITE_DCHECK_LE(params.input1_offset, max_value);
+  TFLITE_DCHECK_GE(params.input2_offset, -max_value);
+  TFLITE_DCHECK_LE(params.input2_offset, max_value);
+  TFLITE_DCHECK_GE(params.output_offset, -max_value);
+  TFLITE_DCHECK_LE(params.output_offset, max_value);
+}
+
+// Element-wise div that can often be used for inner loop of broadcast Div as
+// well as the non-broadcast Div.
+template <typename T>
+inline void DivElementwise(int size, const ArithmeticParams& params,
+                           const T* input1_data, const T* input2_data,
+                           T* output_data) {
+  DivCheckArithmeticParams<T>(params);
+
+  for (int i = 0; i < size; ++i) {
+    int32_t input1_val = params.input1_offset + input1_data[i];
+    int32_t input2_val = params.input2_offset + input2_data[i];
+    TFLITE_DCHECK_NE(input2_val, 0);
+    if (input2_val < 0) {
+      // Invert signs to avoid a negative input2_val as input2_inv needs to be
+      // positive to be used as multiplier of MultiplyByQuantizedMultiplier.
+      input1_val = -input1_val;
+      input2_val = -input2_val;
+    }
+    int recip_shift;
+    const int32_t input2_inv = GetReciprocal(input2_val, 31, &recip_shift);
+    const int headroom = CountLeadingSignBits(input1_val);
+    const int32_t unscaled_quotient =
+        MultiplyByQuantizedMultiplierGreaterThanOne(input1_val, input2_inv,
+                                                    headroom);
+    const int total_shift = params.output_shift - recip_shift - headroom;
+    const int32_t unclamped_result =
+        params.output_offset +
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            unscaled_quotient, params.output_multiplier, total_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);
+  }
+}
+
+inline void Div(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);
+
+  DivElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void Div(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);
+
+  DivElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void Div(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);
+
+  DivElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+template <typename T, int N = 5>
+inline void BroadcastDivSlowQuantized(
+    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) {
+  DivCheckArithmeticParams<T>(params);
+
+  auto op = [&params](T a, T b) {
+    int32_t input1_val = params.input1_offset + a;
+    int32_t input2_val = params.input2_offset + b;
+    TFLITE_DCHECK_NE(input2_val, 0);
+    if (input2_val < 0) {
+      input1_val = -input1_val;
+      input2_val = -input2_val;
+    }
+    int recip_shift;
+    const int32_t input2_inv = GetReciprocal(input2_val, 31, &recip_shift);
+    const int headroom = CountLeadingSignBits(input1_val);
+    const int32_t unscaled_quotient =
+        MultiplyByQuantizedMultiplierGreaterThanOne(input1_val, input2_inv,
+                                                    headroom);
+    const int total_shift = params.output_shift - recip_shift - headroom;
+    const int32_t unclamped_result =
+        params.output_offset +
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            unscaled_quotient, params.output_multiplier, total_shift);
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, unclamped_result));
+    return static_cast<T>(clamped_output);
+  };
+
+  BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                          unextended_input2_shape, input2_data,
+                          unextended_output_shape, output_data, op);
+}
+
+template <int N = 5>
+inline void BroadcastDivSlow(const ArithmeticParams& params,
+                             const RuntimeShape& unextended_input1_shape,
+                             const uint8_t* input1_data,
+                             const RuntimeShape& unextended_input2_shape,
+                             const uint8_t* input2_data,
+                             const RuntimeShape& unextended_output_shape,
+                             uint8_t* output_data) {
+  BroadcastDivSlowQuantized<uint8_t, N>(
+      params, unextended_input1_shape, input1_data, unextended_input2_shape,
+      input2_data, unextended_output_shape, output_data);
+}
+
+template <int N = 5>
+inline void BroadcastDivSlow(const ArithmeticParams& params,
+                             const RuntimeShape& unextended_input1_shape,
+                             const int8_t* input1_data,
+                             const RuntimeShape& unextended_input2_shape,
+                             const int8_t* input2_data,
+                             const RuntimeShape& unextended_output_shape,
+                             int8_t* output_data) {
+  BroadcastDivSlowQuantized<int8_t, N>(
+      params, unextended_input1_shape, input1_data, unextended_input2_shape,
+      input2_data, unextended_output_shape, output_data);
+}
+
+template <int N = 5>
+inline void BroadcastDivSlow(const ArithmeticParams& params,
+                             const RuntimeShape& unextended_input1_shape,
+                             const int16_t* input1_data,
+                             const RuntimeShape& unextended_input2_shape,
+                             const int16_t* input2_data,
+                             const RuntimeShape& unextended_output_shape,
+                             int16_t* output_data) {
+  BroadcastDivSlowQuantized<int16_t, N>(
+      params, unextended_input1_shape, input1_data, unextended_input2_shape,
+      input2_data, unextended_output_shape, output_data);
+}
+
+template <typename T, int N = 5>
+void BroadcastDivSlow(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);
+
+  auto op = [output_activation_min, output_activation_max](T a, T b) {
+    return ActivationFunctionWithMinMax(a / b, output_activation_min,
+                                        output_activation_max);
+  };
+
+  BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                          unextended_input2_shape, input2_data,
+                          unextended_output_shape, output_data, op);
+}
+
+template <typename T>
+inline void Div(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 =
+      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], output_activation_min,
+        output_activation_max);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DIV_H_

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_

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_

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_

tensorflow/lite/kernels/internal/reference/floor.h

@@ -0,0 +1,41 @@
+/* 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 {
+
+template <typename T>
+inline void Floor(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++) {
+    int offset = i;
+    output_data[offset] =
+        static_cast<T>(std::floor(static_cast<float>(input_data[offset])));
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_

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_

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_

tensorflow/lite/kernels/internal/reference/fully_connected.h

@@ -0,0 +1,432 @@
+/* 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 <algorithm>
+#include <cmath>
+#include <cstdint>
+
+#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);
+    }
+  }
+}
+
+// This implementation receives the scales in float and performs requant in
+// float to avoid loss of precision.
+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,
+    float input_scale, float output_scale, float filter_scale,
+    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_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];
+      }
+      const double effective_output_scale = static_cast<double>(input_scale) *
+                                            static_cast<double>(filter_scale) /
+                                            static_cast<double>(output_scale);
+      int32_t acc_scaled = static_cast<int32_t>(
+          round(static_cast<double>(acc) * effective_output_scale));
+      acc_scaled += output_offset;
+      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<uint8_t>(acc_scaled);
+    }
+  }
+}
+
+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;
+    }
+  }
+}
+
+// This implementation receives the scales in float and performs requant in
+// float to avoid loss of precision.
+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,
+    float input_scale, float output_scale, float filter_scale,
+    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_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;
+      }
+      const double effective_output_scale = static_cast<double>(input_scale) *
+                                            static_cast<double>(filter_scale) /
+                                            static_cast<double>(output_scale);
+      int32_t acc_scaled = static_cast<int32_t>(
+          round(static_cast<double>(accum) * effective_output_scale));
+      // Saturate, cast to int16_t, and store to output array.
+      acc_scaled = std::max(acc_scaled, output_activation_min - output_offset);
+      acc_scaled = std::min(acc_scaled, output_activation_max - output_offset);
+      acc_scaled += output_offset;
+      output_data[out_c + output_depth * b] = acc_scaled;
+    }
+  }
+}
+
+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_

tensorflow/lite/kernels/internal/reference/gather.h

@@ -0,0 +1,120 @@
+/* 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_GATHER_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_GATHER_H_
+
+#include <cstdint>
+#include <cstring>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/core/c/c_api_types.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/runtime_shape.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T, typename CoordsT = int32_t>
+inline TfLiteStatus Gather(const tflite::GatherParams& op_params,
+                           const RuntimeShape& input_shape, const T* input_data,
+                           const RuntimeShape& coords_shape,
+                           const CoordsT* coords_data,
+                           const RuntimeShape& output_shape, T* output_data,
+                           bool int4_input = false) {
+  ruy::profiler::ScopeLabel label("Gather");
+  int axis = op_params.axis;
+  if (axis < 0) {
+    axis += input_shape.DimensionsCount();
+  }
+  TFLITE_DCHECK_GE(axis, 0);
+  TFLITE_DCHECK_LT(axis, input_shape.DimensionsCount());
+
+  int batch_dims = op_params.batch_dims;
+  if (batch_dims < 0) {
+    batch_dims += coords_shape.DimensionsCount();
+  }
+  TFLITE_DCHECK_GE(batch_dims, 0);
+  TFLITE_DCHECK_LT(batch_dims, input_shape.DimensionsCount());
+  TFLITE_DCHECK_LE(batch_dims, coords_shape.DimensionsCount());
+  TFLITE_DCHECK_GE(axis, batch_dims);
+  for (int i = 0; i < batch_dims; ++i) {
+    TFLITE_DCHECK_EQ(input_shape.Dims(i), coords_shape.Dims(i));
+  }
+
+  const int axis_size = input_shape.Dims(axis);
+
+  int batch_size = 1;
+  for (int i = 0; i < batch_dims; ++i) {
+    batch_size *= input_shape.Dims(i);
+  }
+
+  int outer_size = 1;
+  for (int i = batch_dims; i < axis; ++i) {
+    outer_size *= input_shape.Dims(i);
+  }
+
+  int inner_size = 1;
+  for (int i = axis + 1; i < input_shape.DimensionsCount(); ++i) {
+    inner_size *= input_shape.Dims(i);
+  }
+
+  int input_flat_size = input_shape.FlatSize();
+  int output_flat_size = output_shape.FlatSize();
+
+  if (int4_input) {
+    // TODO(b/298210669) It doesn't handle the case when sizes are not
+    // divisible by 2.
+    TFLITE_DCHECK_EQ(inner_size % 2, 0);
+    inner_size /= 2;
+    TFLITE_DCHECK_EQ(input_flat_size % 2, 0);
+    input_flat_size /= 2;
+    TFLITE_DCHECK_EQ(output_flat_size % 2, 0);
+    output_flat_size /= 2;
+  }
+
+  int coord_size = 1;
+  for (int i = batch_dims; i < coords_shape.DimensionsCount(); ++i) {
+    coord_size *= coords_shape.Dims(i);
+  }
+
+  for (int batch = 0; batch < batch_size; ++batch) {
+    for (int outer = 0; outer < outer_size; ++outer) {
+      for (int i = 0; i < coord_size; ++i) {
+        // TODO(rsun): replace memcpy with a for loop
+        const int64_t coord = coords_data[batch * coord_size + i];
+        if (coord < 0 || coord >= axis_size) {
+          return kTfLiteError;
+        }
+        const int64_t from_pos =
+            (((batch * outer_size) + outer) * axis_size + coord) * inner_size;
+        TFLITE_DCHECK(from_pos >= 0);
+        TFLITE_DCHECK(from_pos + inner_size <= input_flat_size);
+        const int64_t to_pos =
+            (((batch * outer_size) + outer) * coord_size + i) * inner_size;
+        TFLITE_DCHECK(to_pos >= 0);
+        TFLITE_DCHECK(to_pos + inner_size <= output_flat_size);
+        std::memcpy(&output_data[to_pos], &input_data[from_pos],
+                    sizeof(T) * inner_size);
+      }
+    }
+  }
+  return kTfLiteOk;
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_GATHER_H_

tensorflow/lite/kernels/internal/reference/gelu.h

@@ -0,0 +1,92 @@
+/* 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_GELU_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_GELU_H_
+
+#include <cmath>
+#include <cstdint>
+#include <functional>
+
+#include "Eigen/Core"  // from @eigen_archive
+#include "unsupported/Eigen/CXX11/Tensor"  // from @eigen_archive
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/constants.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+namespace gelu_internal {
+
+constexpr float kSqrt2dPi = M_2_SQRTPI * M_SQRT1_2;  // sqrt( 2 / pi )
+
+}  // namespace gelu_internal
+
+// Plain implementations for GELU. Used for populating lookup table.
+inline float GeluTransform(float in) {
+  // Note: 0.5 * x * ( 1 + erf( x / sqrt( 2 ) ) ) is commonly used, but cause
+  // catastropic cancellation for large negative inputs. Rewriting the
+  // expression via erfc avoids the numerical stability issues.
+  return 0.5f * in * std::erfc(in * static_cast<float>(-M_SQRT1_2));
+}
+
+inline float GeluTransformApproximate(float in) {
+  // 0.5 * x * ( 1 + tanh( sqrt( 2 / pi ) * ( x + 0.044715 * x^3 ) ) )
+  return 0.5f * in *
+         (1.f + std::tanh(gelu_internal::kSqrt2dPi *
+                          // Note: Avoid std::pow for integer exponents
+                          // as it leads to much slower performance.
+                          (in + 0.044715f * in * in * in)));
+}
+
+template <typename T>
+inline void Gelu(const RuntimeShape& input_shape, const T* input_data,
+                 bool approximate, const RuntimeShape& output_shape,
+                 T* output_data) {
+  using VectorType = Eigen::VectorX<T>;
+  auto input_map = VectorType::Map(input_data, input_shape.FlatSize());
+  auto output_map = VectorType::Map(output_data, output_shape.FlatSize());
+
+  if (approximate) {
+    // 0.5 * x * ( 1 + tanh( sqrt( 2 / pi ) * ( x + 0.044715 * x^3 ) ) )
+    output_map.array() = static_cast<T>(0.5) * input_map.array() *
+                         (static_cast<T>(1) +
+                          (static_cast<T>(gelu_internal::kSqrt2dPi) *
+                           (input_map.array() + static_cast<T>(0.044715) *
+                                                    input_map.array().cube()))
+                              .tanh());
+  } else {
+    // Note: 0.5 * x * ( 1 + erf( x / sqrt( 2 ) ) ) is commonly used, but cause
+    // catastropic cancellation for large negative inputs. Rewriting the
+    // expression via erfc avoids the numerical stability issues.
+    output_map.array() =
+        static_cast<T>(0.5) * input_map.array() *
+        (input_map.array() * static_cast<T>(-M_SQRT1_2)).erfc();
+  }
+}
+
+// LookupTableInt16 is a specialized function for int16_t inputs and outputs.
+// It internally calls LUTLookup for table access.
+inline void LookupTableInt16(const int16_t* input_data, int num_elements,
+                             const int16_t* lut, int16_t* output_data) {
+  for (int i = 0; i < num_elements; ++i) {
+    output_data[i] = LUTLookup(input_data[i], lut);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_GELU_H_

tensorflow/lite/kernels/internal/reference/hard_swish.h

@@ -0,0 +1,168 @@
+/* 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_HARD_SWISH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_HARD_SWISH_H_
+
+#include <algorithm>
+
+#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) {
+  int64_t result = static_cast<int64_t>(value) * (1 << amount);
+  result = std::min<int64_t>(result, std::numeric_limits<int16_t>::max());
+  result = std::max<int64_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_HARD_SWISH_H_

tensorflow/lite/kernels/internal/reference/integer_ops/README.md

@@ -0,0 +1,8 @@
+This directory contains reference implementations for int8 fully integer kernels.
+
+Weight filters of convs are expected to be symmetric per-channel quantized in
+the range [-127, 127].
+Inputs/activations are expected to be asymmetric per-layer quantized in the
+range [-128, 127].
+
+THESE ARE EXPERIMENTAL AND PRONE TO CHANGE.

tensorflow/lite/kernels/internal/reference/integer_ops/add.h

@@ -0,0 +1,174 @@
+/* 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 <algorithm>
+#include <cstddef>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.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());
+}
+
+// TODO: b/270589088 - move to a more appropriate file (b/270589088#comment2)
+template <typename T>
+void BroadcastInput1(int size, const ArithmeticParams& params,
+                     const T* input1_data, const T* input2_data, T* output_data,
+                     void (*check_arithmetic_params)(const ArithmeticParams&),
+                     T (*binary_func)(T, T, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[0], input2_data[i], params);
+  }
+}
+
+template <typename T>
+void BroadcastInput2(int size, const ArithmeticParams& params,
+                     const T* input1_data, const T* input2_data, T* output_data,
+                     void (*check_arithmetic_params)(const ArithmeticParams&),
+                     T (*binary_func)(T, T, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[i], input2_data[0], params);
+  }
+}
+
+// TODO: b/270589088 - move to a more appropriate file (b/270589088#comment2)
+template <typename T>
+void ElementWise(int size, const ArithmeticParams& params, const T* input1_data,
+                 const T* input2_data, T* output_data,
+                 void (*check_arithmetic_params)(const ArithmeticParams&),
+                 T (*binary_func)(T, T, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[i], input2_data[i], params);
+  }
+}
+
+// TODO: b/270589088 - move to a more appropriate file. (b/270589088#comment2)
+template <typename T>
+void BroadcastBinaryFunction6DSlow(
+    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,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    T (*binary_func)(T, T, const ArithmeticParams&)) {
+  check_arithmetic_params(params);
+  auto op = [&params, binary_func](T a, T b) {
+    return binary_func(a, b, params);
+  };
+  reference_ops::BroadcastBinaryOpSimple(input1_shape, input1_data,
+                                         input2_shape, input2_data,
+                                         output_shape, output_data, op);
+}
+
+template <typename T>
+void BroadcastBinaryFunction4DSlow(
+    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,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    T (*binary_func)(T, T, const ArithmeticParams&)) {
+  BroadcastBinaryFunction6DSlow(params, input1_shape, input1_data, input2_shape,
+                                input2_data, output_shape, output_data,
+                                check_arithmetic_params, binary_func);
+}
+
+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 BroadcastAdd6DSlow(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) {
+  BroadcastBinaryFunction6DSlow(params, input1_shape, input1_data, input2_shape,
+                                input2_data, output_shape, output_data,
+                                CheckArithmeticParams, AddFunc);
+}
+
+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) {
+  BroadcastBinaryFunction6DSlow(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_

tensorflow/lite/kernels/internal/reference/integer_ops/conv.h

@@ -0,0 +1,241 @@
+/* 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 <algorithm>
+
+#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 = input_shape.Dims(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 filter_input_depth = filter_shape.Dims(3);
+  const int groups = input_depth / filter_input_depth;
+  TFLITE_DCHECK_NE(groups, 0);
+  TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
+  const int filters_per_group = output_depth / groups;
+  TFLITE_DCHECK_NE(filters_per_group, 0);
+  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) {
+          auto group = out_channel / filters_per_group;
+          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 < filter_input_depth;
+                   ++in_channel) {
+                int32_t input_val =
+                    input_data[Offset(input_shape, batch, in_y, in_x,
+                                      in_channel + group * filter_input_depth)];
+                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
+template <typename AccumScalar>
+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 AccumScalar* 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 = input_shape.Dims(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 filter_input_depth = filter_shape.Dims(3);
+  const int groups = input_depth / filter_input_depth;
+  TFLITE_DCHECK_EQ(input_depth % filter_input_depth, 0);
+  const int filters_per_group = output_depth / groups;
+  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) {
+          auto group = out_channel / filters_per_group;
+          AccumScalar 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 < filter_input_depth;
+                   ++in_channel) {
+                int32_t input_val =
+                    input_data[Offset(input_shape, batch, in_y, in_x,
+                                      in_channel + group * filter_input_depth)];
+                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_

tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h

@@ -0,0 +1,291 @@
+/* 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 <algorithm>
+
+#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_

tensorflow/lite/kernels/internal/reference/integer_ops/dequantize.h

@@ -0,0 +1,43 @@
+/* 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_DEQUANTIZE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEQUANTIZE_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+template <typename T>
+inline void Dequantize(const tflite::DequantizationParams& op_params,
+                       const RuntimeShape& input_shape, const T* input_data,
+                       const RuntimeShape& output_shape, float* 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);
+
+  for (int i = 0; i < flat_size; i++) {
+    const int32_t val = static_cast<int32_t>(input_data[i]);
+    const float result = static_cast<float>(scale * (val - zero_point));
+    output_data[i] = result;
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEQUANTIZE_H_

tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h

@@ -0,0 +1,238 @@
+/* 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 <algorithm>
+#include <cmath>
+#include <cstdint>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+// For per-channel functions, since it is defined in quantization spec that
+// weights are symmetric
+// (https://www.tensorflow.org/lite/performance/quantization_spec#symmetric_vs_asymmetric),
+// zero_point (params.weights_offset) is always 0.
+// However, for per-tensor functions, params.weights_offset is still applied for
+// backward compatibility.
+template <typename InputType, typename WeightType, typename OutputType,
+          typename BiasType>
+void FullyConnectedPerChannel(
+    const FullyConnectedParams& params, const int32_t* output_multiplier,
+    const int* output_shift, const RuntimeShape& input_shape,
+    const InputType* input_data, const RuntimeShape& filter_shape,
+    const WeightType* filter_data, const RuntimeShape& bias_shape,
+    const BiasType* bias_data, const RuntimeShape& output_shape,
+    OutputType* output_data) {
+  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;
+  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) {
+      BiasType 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 * (input_val + input_offset);
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+      int32_t acc_scaled = MultiplyByQuantizedMultiplier(
+          acc, output_multiplier[out_c], output_shift[out_c]);
+      acc_scaled += output_offset;
+      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<OutputType>(acc_scaled);
+    }
+  }
+}
+
+// This implementation receives the scales in float and performs requant in
+// float to avoid loss of precision.
+template <typename InputType, typename WeightType, typename OutputType,
+          typename BiasType>
+void FullyConnectedPerChannel(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const InputType* input_data, const RuntimeShape& filter_shape,
+    const WeightType* filter_data, const RuntimeShape& bias_shape,
+    const BiasType* bias_data, const RuntimeShape& output_shape,
+    float input_scale, float output_scale, const float* filter_scales,
+    OutputType* output_data) {
+  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;
+  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) {
+      BiasType 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 * (input_val + input_offset);
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+
+      const float scale = filter_scales[out_c];
+      const double filter_scale = static_cast<double>(scale);
+      const double effective_output_scale = static_cast<double>(input_scale) *
+                                            filter_scale /
+                                            static_cast<double>(output_scale);
+      int32_t acc_scaled = static_cast<int32_t>(
+          round(static_cast<double>(acc) * effective_output_scale));
+
+      acc_scaled += output_offset;
+      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<OutputType>(acc_scaled);
+    }
+  }
+}
+
+template <typename InputType, typename WeightType, typename OutputType,
+          typename BiasType>
+void FullyConnected(const FullyConnectedParams& params,
+                    const RuntimeShape& input_shape,
+                    const InputType* input_data,
+                    const RuntimeShape& filter_shape,
+                    const WeightType* filter_data,
+                    const RuntimeShape& bias_shape, const BiasType* bias_data,
+                    const RuntimeShape& output_shape, OutputType* 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);
+  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) {
+      BiasType 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];
+      }
+      int32_t acc_scaled =
+          MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc_scaled += output_offset;
+      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<OutputType>(acc_scaled);
+    }
+  }
+}
+
+// This implementation receives the scales in float and performs requant in
+// float to avoid loss of precision.
+template <typename InputType, typename WeightType, typename OutputType,
+          typename BiasType>
+void FullyConnected(const FullyConnectedParams& params,
+                    const RuntimeShape& input_shape,
+                    const InputType* input_data,
+                    const RuntimeShape& filter_shape,
+                    const WeightType* filter_data,
+                    const RuntimeShape& bias_shape, const BiasType* bias_data,
+                    const RuntimeShape& output_shape, float input_scale,
+                    float output_scale, float filter_scale,
+                    OutputType* 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_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) {
+      BiasType 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];
+      }
+      const double effective_output_scale = static_cast<double>(input_scale) *
+                                            static_cast<double>(filter_scale) /
+                                            static_cast<double>(output_scale);
+      int32_t acc_scaled = static_cast<int32_t>(
+          round(static_cast<double>(acc) * effective_output_scale));
+      acc_scaled += output_offset;
+      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<OutputType>(acc_scaled);
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_

tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h

@@ -0,0 +1,67 @@
+/* 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 <algorithm>
+
+#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_

tensorflow/lite/kernels/internal/reference/integer_ops/log_softmax.h

@@ -0,0 +1,113 @@
+/* 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_LOG_SOFTMAX_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOG_SOFTMAX_H_
+
+#include <algorithm>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void LogSoftmax(int32_t input_multiplier, int32_t input_shift,
+                       int32_t reverse_multiplier, int32_t reverse_shift,
+                       int32_t diff_min, 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();
+  static constexpr int32_t kMinInt32 = std::numeric_limits<int32_t>::min();
+
+  // [-16, 0] is mapped to [-128, 127] with 1/16 as scale and 127 as zero
+  // point. This nudges the output to [-255/16, 0].
+  static constexpr int32_t kOutputZeroPoint = 127;
+
+  // 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 (int outer_index = 0; outer_index < outer_size; ++outer_index) {
+    int8_t max_in_row = kMinInt8;
+    for (int 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 (int 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_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(
+        diff_min - 1,
+        MultiplyByQuantizedMultiplier(shifted_log_sum_of_exps_in_q5,
+                                      reverse_multiplier, -reverse_shift));
+
+    for (int 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_shift);
+
+        // Rescale and downcast.
+        int32_t output_in_q27 =
+            gemmlowp::RoundingDivideByPOT(
+                (input_diff_in_q5 - log_sum_of_exps_in_q5),
+                31 - kInputIntegerBits - kOutputIntegerBits) +
+            kOutputZeroPoint;
+
+        output_in_q27 =
+            std::max(std::min(output_in_q27, static_cast<int32_t>(kMaxInt8)),
+                     static_cast<int32_t>(kMinInt8));
+        output_data[outer_index * depth + inner_index] =
+            static_cast<int8_t>(output_in_q27);
+      } else {
+        output_data[outer_index * depth + inner_index] = kMinInt8;
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOG_SOFTMAX_H_

tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h

@@ -0,0 +1,121 @@
+/* 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 <algorithm>
+#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_

tensorflow/lite/kernels/internal/reference/integer_ops/lut.h

@@ -0,0 +1,34 @@
+/* Copyright 2022 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_LUT_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LUT_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+template <typename InputT, typename OutputT>
+inline void LookupTable(const InputT* input_data, int num_elements,
+                        const OutputT* lut, OutputT* output_data) {
+  for (int i = 0; i < num_elements; ++i) {
+    output_data[i] = LUTLookup(input_data[i], lut);
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LUT_H_

tensorflow/lite/kernels/internal/reference/integer_ops/mean.h

@@ -0,0 +1,18 @@
+/* Copyright 2023 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_
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_

tensorflow/lite/kernels/internal/reference/integer_ops/mul.h

@@ -0,0 +1,129 @@
+/* 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 <algorithm>
+
+#include "fixedpoint/fixedpoint.h"
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+// Maximum dimension supported by the broadcast mul operation.
+constexpr int kMaxMulBroadcastDim = 6;
+
+template <typename InputType, typename OutputType>
+void MulElementwise(int size, const ArithmeticParams& params,
+                    const InputType* input1_data, const InputType* input2_data,
+                    OutputType* 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<OutputType>(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 BroadcastMul6DSlow(
+    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("BroadcastMul6DSlow");
+  auto op = [&params](T a, T b) {
+    const int32_t input1_val = params.input1_offset + a;
+    const int32_t input2_val = params.input2_offset + b;
+    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));
+    return static_cast<T>(clamped_output);
+  };
+  reference_ops::BroadcastBinaryOpSimple(input1_shape, input1_data,
+                                         input2_shape, input2_data,
+                                         output_shape, output_data, op);
+}
+
+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) {
+  BroadcastMul6DSlow(params, input1_shape, input1_data, input2_shape,
+                     input2_data, output_shape, output_data);
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_

tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h

@@ -0,0 +1,264 @@
+/* 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 <algorithm>
+#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_

tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h

@@ -0,0 +1,117 @@
+/* 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 <algorithm>
+#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_

tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h

@@ -0,0 +1,224 @@
+/* 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 <algorithm>
+
+#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 = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_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, int32 or int64 accumulator
+template <typename Scalar>
+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 Scalar* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data, const RuntimeShape& im2col_shape, int8_t* im2col_data,
+    Scalar* 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 = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_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(Scalar));
+
+  // 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) {
+          Scalar 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_

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/core/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_

tensorflow/lite/kernels/internal/reference/leaky_relu.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_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();
+
+  // Extract the sign and create a safely positive multiplier outside the loop.
+  // This supports negative alpha values (matching float execution behavior)
+  // while preventing assertion failures, as MultiplyByQuantizedMultiplier
+  // strictly requires a non-negative multiplier.
+  const bool is_alpha_negative = params.output_multiplier_alpha < 0;
+  const int32_t safe_alpha_multiplier = is_alpha_negative
+                                            ? -params.output_multiplier_alpha
+                                            : params.output_multiplier_alpha;
+
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input_value = input_data[i] - params.input_offset;
+
+    int32_t unclamped_output = params.output_offset;
+    if (input_value >= 0) {
+      unclamped_output += MultiplyByQuantizedMultiplier(
+          input_value, params.output_multiplier_identity,
+          params.output_shift_identity);
+    } else {
+      int32_t scaled_alpha_value = MultiplyByQuantizedMultiplier(
+          input_value, safe_alpha_multiplier, params.output_shift_alpha);
+      unclamped_output +=
+          is_alpha_negative ? -scaled_alpha_value : scaled_alpha_value;
+    }
+
+    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_

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_

tensorflow/lite/kernels/internal/reference/logistic.h

@@ -0,0 +1,133 @@
+/* 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 {
+
+template <typename T>
+inline void Logistic(const RuntimeShape& input_shape, const T* input_data,
+                     const RuntimeShape& output_shape, T* 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++) {
+    T 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] = static_cast<T>(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_

tensorflow/lite/kernels/internal/reference/lstm_cell.h

@@ -0,0 +1,422 @@
+/* Copyright 2022 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_LSTM_CELL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LSTM_CELL_H_
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/concatenation.h"
+#include "tensorflow/lite/kernels/internal/reference/fully_connected.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void LstmCell(
+    const LstmCellParams& params, const RuntimeShape& unextended_input_shape,
+    const float* input_data, const RuntimeShape& unextended_prev_activ_shape,
+    const float* prev_activ_data, const RuntimeShape& weights_shape,
+    const float* weights_data, const RuntimeShape& unextended_bias_shape,
+    const float* bias_data, const RuntimeShape& unextended_prev_state_shape,
+    const float* prev_state_data,
+    const RuntimeShape& unextended_output_state_shape, float* output_state_data,
+    const RuntimeShape& unextended_output_activ_shape, float* output_activ_data,
+    const RuntimeShape& unextended_concat_temp_shape, float* concat_temp_data,
+    const RuntimeShape& unextended_activ_temp_shape, float* activ_temp_data) {
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_prev_activ_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_bias_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_prev_state_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_state_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_activ_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_concat_temp_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_activ_temp_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape prev_activ_shape =
+      RuntimeShape::ExtendedShape(4, unextended_prev_activ_shape);
+  const RuntimeShape bias_shape =
+      RuntimeShape::ExtendedShape(4, unextended_bias_shape);
+  const RuntimeShape prev_state_shape =
+      RuntimeShape::ExtendedShape(4, unextended_prev_state_shape);
+  const RuntimeShape output_state_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_state_shape);
+  const RuntimeShape output_activ_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_activ_shape);
+  const RuntimeShape concat_temp_shape =
+      RuntimeShape::ExtendedShape(4, unextended_concat_temp_shape);
+  const RuntimeShape activ_temp_shape =
+      RuntimeShape::ExtendedShape(4, unextended_activ_temp_shape);
+  TFLITE_DCHECK_GE(weights_shape.DimensionsCount(), 2);
+
+  const int weights_dim_count = weights_shape.DimensionsCount();
+  const int batches =
+      MatchingDim(input_shape, 0, prev_activ_shape, 0, prev_state_shape, 0,
+                  output_state_shape, 0, output_activ_shape, 0);
+  const int height =
+      MatchingDim(input_shape, 1, prev_activ_shape, 1, prev_state_shape, 1,
+                  output_state_shape, 1, output_activ_shape, 1);
+  const int width =
+      MatchingDim(input_shape, 2, prev_activ_shape, 2, prev_state_shape, 2,
+                  output_state_shape, 2, output_activ_shape, 2);
+  const int input_depth = input_shape.Dims(3);
+  const int prev_activ_depth = prev_activ_shape.Dims(3);
+  const int total_input_depth = prev_activ_depth + input_depth;
+  TFLITE_DCHECK_EQ(weights_shape.Dims(weights_dim_count - 1),
+                   total_input_depth);
+  TFLITE_DCHECK_EQ(FlatSizeSkipDim(bias_shape, 3), 1);
+  const int intern_activ_depth =
+      MatchingDim(weights_shape, weights_dim_count - 2, bias_shape, 3);
+  TFLITE_DCHECK_EQ(weights_shape.FlatSize(),
+                   intern_activ_depth * total_input_depth);
+  TFLITE_DCHECK_EQ(intern_activ_depth % 4, 0);
+  const int output_depth =
+      MatchingDim(prev_state_shape, 3, prev_activ_shape, 3, output_state_shape,
+                  3, output_activ_shape, 3);
+  TFLITE_DCHECK_EQ(output_depth, intern_activ_depth / 4);
+
+  // Concatenate prev_activ and input data together
+  float const* concat_input_arrays_data[2] = {input_data, prev_activ_data};
+  const RuntimeShape* concat_input_arrays_shapes[2] = {&input_shape,
+                                                       &prev_activ_shape};
+  tflite::ConcatenationParams concat_params;
+  concat_params.axis = 3;
+  concat_params.inputs_count = 2;
+  Concatenation(concat_params, concat_input_arrays_shapes,
+                concat_input_arrays_data, concat_temp_shape, concat_temp_data);
+
+  // Fully connected
+  tflite::FullyConnectedParams fc_params;
+  fc_params.float_activation_min = std::numeric_limits<float>::lowest();
+  fc_params.float_activation_max = std::numeric_limits<float>::max();
+  FullyConnected(fc_params, concat_temp_shape, concat_temp_data, weights_shape,
+                 weights_data, bias_shape, bias_data, activ_temp_shape,
+                 activ_temp_data);
+
+  // Memory state update (the LSTM "guts")
+  for (int b = 0; b < batches; ++b) {
+    for (int w = 0; w < width; ++w) {
+      for (int h = 0; h < height; ++h) {
+        for (int c = 0; c < output_depth; ++c) {
+          const float input_gate =
+              1.f /
+              (1.f + std::exp(-activ_temp_data[Offset(activ_temp_shape, b, h, w,
+                                                      0 * output_depth + c)]));
+          const float new_input = std::tanh(activ_temp_data[Offset(
+              activ_temp_shape, b, h, w, 1 * output_depth + c)]);
+          const float forget_gate =
+              1.f /
+              (1.f + std::exp(-activ_temp_data[Offset(activ_temp_shape, b, h, w,
+                                                      2 * output_depth + c)]));
+          const float output_gate =
+              1.f /
+              (1.f + std::exp(-activ_temp_data[Offset(activ_temp_shape, b, h, w,
+                                                      3 * output_depth + c)]));
+          const float new_state =
+              input_gate * new_input +
+              forget_gate *
+                  prev_state_data[Offset(prev_state_shape, b, h, w, c)];
+          output_state_data[Offset(output_state_shape, b, h, w, c)] = new_state;
+          output_activ_data[Offset(output_activ_shape, b, h, w, c)] =
+              output_gate * std::tanh(new_state);
+        }
+      }
+    }
+  }
+}
+
+// Quantized LSTM cell implementation.
+// The quantization of the input, output arrays is as follows:
+//  - The input activations are quantized as uint8 on the interval
+//    [-1, 127/128].
+//    The rationale for that is that is the natural interval for output
+//    activations (see next point) and these need to be concatenated together.
+//    We could accommodate different ranges by re-scaling, but we empirically
+//    found that setting the input activations range to be [-1, 127/128] in the
+//    first place, removing the need for re-scaling, greatly improves accuracy.
+//  - The output activations are quantized as uint8 on the interval
+//    [-1, 127/128].
+//    The rationale for that is that the definition of a LSTM cell makes them
+//    intrinsically constrained in [-1, 1]; tweaking that to [-1, 127/128]
+//    makes for simpler, more accurate fixed-point arithmetic.
+//  - The output-at-previous-timestep state array is obviously quantized as
+//    the output activations.
+//  - The internal LSTM memory (not the output-at-previous-timestep, the other
+//    internal state array) is int16-quantized and may use any power-of-two,
+//    symmetric range i.e. [-2^N, 2^N * 32767/32768] for any N, which we call
+//    StateIntegerBits below, see the below discussion of that template
+//    parameter ("The StateIntegerBits template parameter").
+//  - The output of the internal fully-connected node is int16-quantized
+//    on the interval [-8, 8 * 32767/32768], the rationale for which is
+//    explained just below ("Why [-8, 8] for fully-connected output?").
+//
+//
+// === The StateIntegerBits template parameter ===
+//
+// The StateIntegerBits template parameter controls the fixed-point format used
+// to represent the internal memory of the LSTM cell (not the
+// output-at-previous-timestep, the other internal state array). It's currently
+// a template parameter so that the model can control that. The most typical
+// value for StateIntegerBits is 4. Other plausible values are anywhere between
+// 3 and 5. We might eventually standardize on a single supported value, e.g. 4,
+// and drop that template parameter. The reason why it can't be a runtime
+// parameter is that this controls the fixed-point format used, i.e. we need to
+// generate actually different code based on it. In particular, we generate code
+// for a fixed-point tanh() implementation for that format, which internally
+// uses a fixed-point exp() implementation, which internally uses a
+// barrel-shifter with a number of steps that depends on StateIntegerBits.
+// Another consequence of that is that a higher value of StateIntegerBits
+// results in a more expensive implementation (more barrel shifter steps
+// needed).
+//
+//
+// === Why [-8, 8] for fully-connected output? ===
+//
+// This array is only fed to Logistic and Tanh functions, for which
+// the quantized implementation will want to use fixed-point arithmetic,
+// requiring a power-of-two representation interval. Thus, we should right
+// away quantize this array to a power-of-two interval; otherwise,
+// implementation will need to rescale that, losing any benefit that a tighter
+// representation interval might otherwise yield, while introducing some
+// numerical error and computational overhead.
+//
+// Now, Logistic and Tanh
+// are nearly constant (nearly equal to their horizontal asymptotes)
+// outside of a small bounded interval around 0:
+//
+//   Logistic(4) = 1 - 1.8e-2     Tanh(4) = 1 - 6.7e-4
+//   Logistic(8) = 1 - 3.4e-4     Tanh(8) = 1 - 2.3e-7
+//   Logistic(16) = 1 - 1.1e-7    Tanh(16) = 1 - 2.5e-14
+//
+// From this, we see that clamping to [-4, 4] would be too inaccurate
+// (the error of 1.8e-2 on Logistic would be felt even in 8bit precision)
+// while clamping to [-16, 16] would make no difference even in float32.
+// However, for a fixed-point implementation in 16-bit integers, using 5
+// integer bits to represent the [-16, 16] range would leave only 11
+// fractional bits, giving an increment of 2^-11 = 4.9e-4 between consecutive
+// representable values. Notice that is higher than the
+// worst-case clamping error with clamping to [-8, 8]: 3.4e-4 for Logistic.
+// Using [-8, 8] thus seems like the better compromise overall, enjoying
+// an increment of 2.4e-4 between representable values and a worst-case
+// clamping error of 3.4e-4, both better than the increment of 4.9e-4 with
+// [-16, 16].
+//
+// Moreover, all other things being equal, it is nice to choose the narrower
+// representation range, as that makes the implementation of fixed-point
+// math functions a little cheaper (each integer bit requires an additional
+// barrel-shifter atep in the implementation of exp(-x)). That is further
+// reason to prefer [-8, 8] over [-16, 16]. The choice of [-16, 16] would make
+// sense for 32-bit float or 32-bit fixed-point quantization, but we are
+// aiming for 16-bit fixed-point quantization of these internal nodes here.
+//
+template <int StateIntegerBits>
+inline void LstmCell(const LstmCellParams& params,
+                     const RuntimeShape& unextended_input_shape,
+                     const uint8_t* input_data_uint8,
+                     const RuntimeShape& unextended_prev_activ_shape,
+                     const uint8_t* prev_activ_data_uint8,
+                     const RuntimeShape& weights_shape,
+                     const uint8_t* weights_data_uint8,
+                     const RuntimeShape& unextended_bias_shape,
+                     const int32_t* bias_data_int32,
+                     const RuntimeShape& unextended_prev_state_shape,
+                     const int16_t* prev_state_data_int16,
+                     const RuntimeShape& unextended_output_state_shape,
+                     int16_t* output_state_data_int16,
+                     const RuntimeShape& unextended_output_activ_shape,
+                     uint8_t* output_activ_data_uint8,
+                     const RuntimeShape& unextended_concat_temp_shape,
+                     uint8_t* concat_temp_data_uint8,
+                     const RuntimeShape& unextended_activ_temp_shape,
+                     int16_t* activ_temp_data_int16, void* gemmlowp_context) {
+  (void)gemmlowp_context;  // only used in optimized code.
+  int32_t weights_zero_point = params.weights_zero_point;
+  int32_t accum_multiplier = params.accum_multiplier;
+  int accum_shift = params.accum_shift;
+  TFLITE_DCHECK_LE(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_prev_activ_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_bias_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_prev_state_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_state_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_activ_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_concat_temp_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_activ_temp_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape prev_activ_shape =
+      RuntimeShape::ExtendedShape(4, unextended_prev_activ_shape);
+  const RuntimeShape bias_shape =
+      RuntimeShape::ExtendedShape(4, unextended_bias_shape);
+  const RuntimeShape prev_state_shape =
+      RuntimeShape::ExtendedShape(4, unextended_prev_state_shape);
+  const RuntimeShape output_state_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_state_shape);
+  const RuntimeShape output_activ_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_activ_shape);
+  const RuntimeShape concat_temp_shape =
+      RuntimeShape::ExtendedShape(4, unextended_concat_temp_shape);
+  const RuntimeShape activ_temp_shape =
+      RuntimeShape::ExtendedShape(4, unextended_activ_temp_shape);
+  TFLITE_DCHECK_GE(weights_shape.DimensionsCount(), 2);
+
+  // Gather dimensions information, and perform consistency checks.
+  const int weights_dim_count = weights_shape.DimensionsCount();
+  const int outer_size = MatchingFlatSizeSkipDim(
+      input_shape, 3, prev_activ_shape, prev_state_shape, output_state_shape,
+      output_activ_shape);
+  const int input_depth = input_shape.Dims(3);
+  const int prev_activ_depth = prev_activ_shape.Dims(3);
+  const int total_input_depth = prev_activ_depth + input_depth;
+  TFLITE_DCHECK_EQ(weights_shape.Dims(weights_dim_count - 1),
+                   total_input_depth);
+  const int intern_activ_depth =
+      MatchingDim(weights_shape, weights_dim_count - 2, bias_shape, 3);
+  TFLITE_DCHECK_EQ(weights_shape.FlatSize(),
+                   intern_activ_depth * total_input_depth);
+  TFLITE_DCHECK_EQ(FlatSizeSkipDim(bias_shape, 3), 1);
+  TFLITE_DCHECK_EQ(intern_activ_depth % 4, 0);
+  const int output_depth =
+      MatchingDim(prev_state_shape, 3, prev_activ_shape, 3, output_state_shape,
+                  3, output_activ_shape, 3);
+  TFLITE_DCHECK_EQ(output_depth, intern_activ_depth / 4);
+  const int fc_batches = FlatSizeSkipDim(activ_temp_shape, 3);
+  const int fc_output_depth =
+      MatchingDim(weights_shape, weights_dim_count - 2, activ_temp_shape, 3);
+  const int fc_accum_depth = total_input_depth;
+  TFLITE_DCHECK_EQ(fc_output_depth, 4 * output_depth);
+
+  // Depth-concatenate prev_activ and input data together.
+  uint8_t const* concat_input_arrays_data[2] = {input_data_uint8,
+                                                prev_activ_data_uint8};
+  const RuntimeShape* concat_input_arrays_shapes[2] = {&input_shape,
+                                                       &prev_activ_shape};
+  tflite::ConcatenationParams concat_params;
+  concat_params.axis = 3;
+  concat_params.inputs_count = 2;
+  Concatenation(concat_params, concat_input_arrays_shapes,
+                concat_input_arrays_data, concat_temp_shape,
+                concat_temp_data_uint8);
+
+  // Implementation of the fully connected node inside the LSTM cell.
+  // The operands are 8-bit integers, the accumulators are internally 32bit
+  // integers, and the output is 16-bit fixed-point with 3 integer bits so
+  // the output range is [-2^3, 2^3] == [-8, 8]. The rationale for that
+  // is explained in the function comment above.
+  for (int b = 0; b < fc_batches; ++b) {
+    for (int out_c = 0; out_c < fc_output_depth; ++out_c) {
+      // Internal accumulation.
+      // Initialize accumulator with the bias-value.
+      int32_t accum = bias_data_int32[out_c];
+      // Accumulation loop.
+      for (int d = 0; d < fc_accum_depth; ++d) {
+        int16_t input_val =
+            concat_temp_data_uint8[b * fc_accum_depth + d] - 128;
+        int16_t weights_val =
+            weights_data_uint8[out_c * fc_accum_depth + d] - weights_zero_point;
+        accum += input_val * weights_val;
+      }
+      // Down-scale the final int32 accumulator to the scale used by our
+      // (16-bit, using 3 integer bits) fixed-point format. The quantized
+      // multiplier and shift here have been pre-computed offline
+      // (e.g. by toco).
+      accum =
+          MultiplyByQuantizedMultiplier(accum, accum_multiplier, accum_shift);
+      // Saturate, cast to int16, and store to the temporary activations array.
+      accum = std::max(-32768, std::min(32767, accum));
+      activ_temp_data_int16[out_c + fc_output_depth * b] = accum;
+    }
+  }
+
+  // Rest of the LSTM cell: tanh and logistic math functions, and some adds
+  // and muls, all done in 16-bit fixed-point.
+  for (int b = 0; b < outer_size; ++b) {
+    for (int c = 0; c < output_depth; ++c) {
+      // Define the fixed-point data types that we will use here. All use
+      // int16 as the underlying integer type i.e. all are 16-bit fixed-point.
+      // They only differ by the number of integral vs. fractional bits,
+      // determining the range of values that they can represent.
+      //
+      // 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].
+      // This is the range of the previous fully-connected node's output,
+      // which is our input here.
+      using F3 = gemmlowp::FixedPoint<std::int16_t, 3>;
+      // FS uses StateIntegerBits integer bits, range [-2^StateIntegerBits,
+      // 2^StateIntegerBits]. It's used to represent the internal state, whose
+      // number of integer bits is currently dictated by the model. See comment
+      // on the StateIntegerBits template parameter above.
+      using FS = gemmlowp::FixedPoint<std::int16_t, StateIntegerBits>;
+      // Implementation of input gate, using fixed-point logistic function.
+      F3 input_gate_input = F3::FromRaw(
+          activ_temp_data_int16[b * fc_output_depth + 0 * output_depth + c]);
+      F0 input_gate_output = gemmlowp::logistic(input_gate_input);
+      // Implementation of input modulation gate, using fixed-point tanh
+      // function.
+      F3 input_modulation_gate_input = F3::FromRaw(
+          activ_temp_data_int16[b * fc_output_depth + 1 * output_depth + c]);
+      F0 input_modulation_gate_output =
+          gemmlowp::tanh(input_modulation_gate_input);
+      // Implementation of forget gate, using fixed-point logistic function.
+      F3 forget_gate_input = F3::FromRaw(
+          activ_temp_data_int16[b * fc_output_depth + 2 * output_depth + c]);
+      F0 forget_gate_output = gemmlowp::logistic(forget_gate_input);
+      // Implementation of output gate, using fixed-point logistic function.
+      F3 output_gate_input = F3::FromRaw(
+          activ_temp_data_int16[b * fc_output_depth + 3 * output_depth + c]);
+      F0 output_gate_output = gemmlowp::logistic(output_gate_input);
+      // Implementation of internal multiplication nodes, still in fixed-point.
+      F0 input_times_input_modulation =
+          input_gate_output * input_modulation_gate_output;
+      FS prev_state = FS::FromRaw(prev_state_data_int16[b * output_depth + c]);
+      FS prev_state_times_forget_state = forget_gate_output * prev_state;
+      // Implementation of internal addition node, saturating.
+      FS new_state = gemmlowp::SaturatingAdd(
+          gemmlowp::Rescale<StateIntegerBits>(input_times_input_modulation),
+          prev_state_times_forget_state);
+      // Implementation of last internal Tanh node, still in fixed-point.
+      // Since a Tanh fixed-point implementation is specialized for a given
+      // number or integer bits, and each specialization can have a substantial
+      // code size, and we already used above a Tanh on an input with 3 integer
+      // bits, and per the table in the above function comment there is no
+      // significant accuracy to be lost by clamping to [-8, +8] for a
+      // 3-integer-bits representation, let us just do that. This helps people
+      // porting this to targets where code footprint must be minimized.
+      F3 new_state_f3 = gemmlowp::Rescale<3>(new_state);
+      F0 output_activ_int16 = output_gate_output * gemmlowp::tanh(new_state_f3);
+      // Store the new internal state back to memory, as 16-bit integers.
+      // Note: here we store the original value with StateIntegerBits, not
+      // the rescaled 3-integer-bits value fed to tanh.
+      output_state_data_int16[b * output_depth + c] = new_state.raw();
+      // Down-scale the output activations to 8-bit integers, saturating,
+      // and store back to memory.
+      int16_t rescaled_output_activ =
+          gemmlowp::RoundingDivideByPOT(output_activ_int16.raw(), 8);
+      int16_t clamped_output_activ = std::max<int16_t>(
+          -128, std::min<int16_t>(127, rescaled_output_activ));
+      output_activ_data_uint8[b * output_depth + c] =
+          128 + clamped_output_activ;
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LSTM_CELL_H_

tensorflow/lite/kernels/internal/reference/maximum_minimum.h

@@ -0,0 +1,50 @@
+/* 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/reference/broadcast_loop.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 {
+    BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                            unextended_input2_shape, input2_data,
+                            unextended_output_shape, output_data, op);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_

tensorflow/lite/kernels/internal/reference/mul.h

@@ -0,0 +1,168 @@
+/* Copyright 2023 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 <algorithm>
+#include <complex>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Maximum dimension supported by the broadcast mul operation.
+constexpr int kMaxMulBroadcastDim = 6;
+
+// 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<T>(
+        input1_data[i] * input2_data[i], output_activation_min,
+        output_activation_max);
+  }
+}
+
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape,
+                const std::complex<float>* input1_data,
+                const RuntimeShape& input2_shape,
+                const std::complex<float>* input2_data,
+                const RuntimeShape& output_shape,
+                std::complex<float>* output_data) {
+  const int flat_size =
+      MatchingExtendedShapeFlatSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = input1_data[i] * input2_data[i];
+  }
+}
+
+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 BroadcastMul6DSlow(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) {
+  auto op = [&params](uint8_t input1_val, uint8_t input2_val) {
+    const int32_t offsetted_input1_val = params.input1_offset + input1_val;
+    const int32_t offsetted_input2_val = params.input2_offset + input2_val;
+    const int32_t unclamped_result =
+        params.output_offset + MultiplyByQuantizedMultiplier(
+                                   offsetted_input1_val * offsetted_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));
+    return static_cast<uint8_t>(clamped_output);
+  };
+  BroadcastBinaryOpSimple(input1_shape, input1_data, input2_shape, input2_data,
+                          output_shape, output_data, op);
+}
+
+template <typename T,
+          // For unquantized mul on small integers, explicitly set to true.
+          bool enable_for_short_integers = false>
+inline typename std::enable_if<
+    !is_small_integer<T>::value || enable_for_short_integers, void>::type
+BroadcastMul6DSlow(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);
+  auto op = [output_activation_min, output_activation_max](T a, T b) {
+    return ActivationFunctionWithMinMax<T>(a * b, output_activation_min,
+                                           output_activation_max);
+  };
+  BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                          unextended_input2_shape, input2_data,
+                          unextended_output_shape, output_data, op);
+}
+
+inline void BroadcastMul6DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& unextended_input1_shape,
+                               const std::complex<float>* input1_data,
+                               const RuntimeShape& unextended_input2_shape,
+                               const std::complex<float>* input2_data,
+                               const RuntimeShape& unextended_output_shape,
+                               std::complex<float>* output_data) {
+  auto op = [](std::complex<float> a, std::complex<float> b) { return a * b; };
+  BroadcastBinaryOpSimple(unextended_input1_shape, input1_data,
+                          unextended_input2_shape, input2_data,
+                          unextended_output_shape, output_data, op);
+}
+
+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) {
+  return BroadcastMul6DSlow(params, input1_shape, input1_data, input2_shape,
+                            input2_data, output_shape, output_data);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_

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_

tensorflow/lite/kernels/internal/reference/non_max_suppression.h

@@ -0,0 +1,193 @@
+/* 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_NON_MAX_SUPPRESSION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NON_MAX_SUPPRESSION_H_
+
+#include <algorithm>
+#include <cmath>
+#include <deque>
+#include <queue>
+
+namespace tflite {
+namespace reference_ops {
+
+// A pair of diagonal corners of the box.
+struct BoxCornerEncoding {
+  float y1;
+  float x1;
+  float y2;
+  float x2;
+};
+
+inline float ComputeIntersectionOverUnion(const float* boxes, const int i,
+                                          const int j) {
+  auto& box_i = reinterpret_cast<const BoxCornerEncoding*>(boxes)[i];
+  auto& box_j = reinterpret_cast<const BoxCornerEncoding*>(boxes)[j];
+  const float box_i_y_min = std::min<float>(box_i.y1, box_i.y2);
+  const float box_i_y_max = std::max<float>(box_i.y1, box_i.y2);
+  const float box_i_x_min = std::min<float>(box_i.x1, box_i.x2);
+  const float box_i_x_max = std::max<float>(box_i.x1, box_i.x2);
+  const float box_j_y_min = std::min<float>(box_j.y1, box_j.y2);
+  const float box_j_y_max = std::max<float>(box_j.y1, box_j.y2);
+  const float box_j_x_min = std::min<float>(box_j.x1, box_j.x2);
+  const float box_j_x_max = std::max<float>(box_j.x1, box_j.x2);
+
+  const float area_i =
+      (box_i_y_max - box_i_y_min) * (box_i_x_max - box_i_x_min);
+  const float area_j =
+      (box_j_y_max - box_j_y_min) * (box_j_x_max - box_j_x_min);
+  if (area_i <= 0 || area_j <= 0) return 0.0;
+  const float intersection_ymax = std::min<float>(box_i_y_max, box_j_y_max);
+  const float intersection_xmax = std::min<float>(box_i_x_max, box_j_x_max);
+  const float intersection_ymin = std::max<float>(box_i_y_min, box_j_y_min);
+  const float intersection_xmin = std::max<float>(box_i_x_min, box_j_x_min);
+  const float intersection_area =
+      std::max<float>(intersection_ymax - intersection_ymin, 0.0) *
+      std::max<float>(intersection_xmax - intersection_xmin, 0.0);
+  return intersection_area / (area_i + area_j - intersection_area);
+}
+
+// Implements (Single-Class) Soft NMS (with Gaussian weighting).
+// Supports functionality of TensorFlow ops NonMaxSuppressionV4 & V5.
+// Reference: "Soft-NMS - Improving Object Detection With One Line of Code"
+//            [Bodla et al, https://arxiv.org/abs/1704.04503]
+// Implementation adapted from the TensorFlow NMS code at
+// tensorflow/core/kernels/non_max_suppression_op.cc.
+//
+// Arguments:
+//  boxes: box encodings in format [y1, x1, y2, x2], shape: [num_boxes, 4]
+//  num_boxes: number of candidates
+//  scores: scores for candidate boxes, in the same order. shape: [num_boxes]
+//  max_output_size: the maximum number of selections.
+//  iou_threshold: Intersection-over-Union (IoU) threshold for NMS
+//  score_threshold: All candidate scores below this value are rejected
+//  soft_nms_sigma: Soft NMS parameter, used for decaying scores
+//
+// Outputs:
+//  selected_indices: all the selected indices. Underlying array must have
+//    length >= max_output_size. Cannot be null.
+//  selected_scores: scores of selected indices. Defer from original value for
+//    Soft NMS. If not null, array must have length >= max_output_size.
+//  num_selected_indices: Number of selections. Only these many elements are
+//    set in selected_indices, selected_scores. Cannot be null.
+//
+// Assumes inputs are valid (for eg, iou_threshold must be >= 0).
+inline void NonMaxSuppression(const float* boxes, const int num_boxes,
+                              const float* scores, const int max_output_size,
+                              const float iou_threshold,
+                              const float score_threshold,
+                              const float soft_nms_sigma, int* selected_indices,
+                              float* selected_scores,
+                              int* num_selected_indices) {
+  struct Candidate {
+    int index;
+    float score;
+    int suppress_begin_index;
+  };
+
+  // Priority queue to hold candidates.
+  auto cmp = [](const Candidate bs_i, const Candidate bs_j) {
+    return bs_i.score < bs_j.score;
+  };
+  std::priority_queue<Candidate, std::deque<Candidate>, decltype(cmp)>
+      candidate_priority_queue(cmp);
+  // Populate queue with candidates above the score threshold.
+  for (int i = 0; i < num_boxes; ++i) {
+    if (scores[i] > score_threshold) {
+      candidate_priority_queue.emplace(Candidate({i, scores[i], 0}));
+    }
+  }
+
+  *num_selected_indices = 0;
+  int num_outputs = std::min(static_cast<int>(candidate_priority_queue.size()),
+                             max_output_size);
+  if (num_outputs == 0) return;
+
+  // NMS loop.
+  float scale = 0;
+  if (soft_nms_sigma > 0.0) {
+    scale = -0.5 / soft_nms_sigma;
+  }
+  while (*num_selected_indices < num_outputs &&
+         !candidate_priority_queue.empty()) {
+    Candidate next_candidate = candidate_priority_queue.top();
+    const float original_score = next_candidate.score;
+    candidate_priority_queue.pop();
+
+    // Overlapping boxes are likely to have similar scores, therefore we
+    // iterate through the previously selected boxes backwards in order to
+    // see if `next_candidate` should be suppressed. We also enforce a property
+    // that a candidate can be suppressed by another candidate no more than
+    // once via `suppress_begin_index` which tracks which previously selected
+    // boxes have already been compared against next_candidate prior to a given
+    // iteration.  These previous selected boxes are then skipped over in the
+    // following loop.
+    bool should_hard_suppress = false;
+    for (int j = *num_selected_indices - 1;
+         j >= next_candidate.suppress_begin_index; --j) {
+      const float iou = ComputeIntersectionOverUnion(
+          boxes, next_candidate.index, selected_indices[j]);
+
+      // First decide whether to perform hard suppression.
+      if (iou >= iou_threshold) {
+        should_hard_suppress = true;
+        break;
+      }
+
+      // Suppress score if NMS sigma > 0.
+      if (soft_nms_sigma > 0.0) {
+        next_candidate.score =
+            next_candidate.score * std::exp(scale * iou * iou);
+      }
+
+      // If score has fallen below score_threshold, it won't be pushed back into
+      // the queue.
+      if (next_candidate.score <= score_threshold) break;
+    }
+    // If `next_candidate.score` has not dropped below `score_threshold`
+    // by this point, then we know that we went through all of the previous
+    // selections and can safely update `suppress_begin_index` to
+    // `selected.size()`. If on the other hand `next_candidate.score`
+    // *has* dropped below the score threshold, then since `suppress_weight`
+    // always returns values in [0, 1], further suppression by items that were
+    // not covered in the above for loop would not have caused the algorithm
+    // to select this item. We thus do the same update to
+    // `suppress_begin_index`, but really, this element will not be added back
+    // into the priority queue.
+    next_candidate.suppress_begin_index = *num_selected_indices;
+
+    if (!should_hard_suppress) {
+      if (next_candidate.score == original_score) {
+        // Suppression has not occurred, so select next_candidate.
+        selected_indices[*num_selected_indices] = next_candidate.index;
+        if (selected_scores) {
+          selected_scores[*num_selected_indices] = next_candidate.score;
+        }
+        ++*num_selected_indices;
+      }
+      if (next_candidate.score > score_threshold) {
+        // Soft suppression might have occurred and current score is still
+        // greater than score_threshold; add next_candidate back onto priority
+        // queue.
+        candidate_priority_queue.push(next_candidate);
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NON_MAX_SUPPRESSION_H_

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_

tensorflow/lite/kernels/internal/reference/pooling.h

@@ -0,0 +1,303 @@
+/* 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 <algorithm>
+
+#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_

tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc

@@ -0,0 +1,819 @@
+/* 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
+
+// LINT.IfChange(portable_symmetric_quantize_floats)
+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)));
+  }
+}
+// LINT.ThenChange(//tensorflow/compiler/mlir/lite/quantization/lite/toco_legacy/portable_tensor_utils.cc:portable_symmetric_quantize_floats)
+
+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 PortableSparseMatrixBatchVectorMultiplyAccumulate1x16(
+    const int8_t* __restrict__ matrix, const int32_t* __restrict__ segments,
+    const int32_t* __restrict__ indices, int m_rows, int m_cols,
+    const int8_t* __restrict__ vector, const int32_t* __restrict__ bias_vector,
+    int n_batch, const int32_t input_offset, const int32_t output_multiplier,
+    const int32_t output_shift, const int32_t* per_channel_scale,
+    const int32_t* per_channel_shift, const int32_t output_offset,
+    const int32_t output_activation_min, const int32_t output_activation_max,
+    int8_t* __restrict__ result) {
+  const int kBlockSize = 16;
+  TFLITE_DCHECK_EQ(m_cols % kBlockSize, 0);
+  for (int batch = 0; batch < n_batch; ++batch) {
+    const int8_t* matrix_ptr = matrix;
+    for (int row = 0; row < m_rows; ++row) {
+      int32_t dot_prod = 0;
+      const int8_t* 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 int8_t* 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++;
+          dot_prod += *matrix_ptr++ * input_offset;
+        }
+      }
+      const int32_t bias_value = bias_vector != nullptr ? bias_vector[row] : 0;
+      dot_prod = MultiplyByQuantizedMultiplier(
+          dot_prod + bias_value,
+          per_channel_scale ? per_channel_scale[row] : output_multiplier,
+          per_channel_shift ? per_channel_shift[row] : output_shift);
+      dot_prod += output_offset;
+      result[batch * m_rows + row] =
+          static_cast<int8_t>(ActivationFunctionWithMinMax(
+              dot_prod, output_activation_min, output_activation_max));
+    }
+  }
+}
+
+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,
+    const float* per_channel_scale) {
+  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
+      float scaling_factor = batch_scaling_factor;
+      if (per_channel_scale) {
+        scaling_factor *= per_channel_scale[row];
+      }
+      result[batch * m_rows + row] += dotprod * 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

tensorflow/lite/kernels/internal/reference/portable_tensor_utils.h

@@ -0,0 +1,336 @@
+/* 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_PORTABLE_TENSOR_UTILS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PORTABLE_TENSOR_UTILS_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 {
+
+// Check if all entries of a vector are zero for float.
+bool IsZeroVector(const float* vector, int v_size) {
+  return PortableIsZeroVector(vector, v_size);
+}
+
+// Check if all entries of a vector are zero for int8_t.
+bool IsZeroVector(const int8_t* vector, int v_size) {
+  return PortableIsZeroVector(vector, v_size);
+}
+
+void SymmetricQuantizeFloats(const float* values, const int size,
+                             int8_t* quantized_values, float* min, float* max,
+                             float* scaling_factor) {
+  PortableSymmetricQuantizeFloats(values, size, quantized_values, min, max,
+                                  scaling_factor);
+}
+
+void SymmetricQuantizeFloats(const float* values, const int size,
+                             int8_t* quantized_values, float min_value,
+                             float max_value, float* scaling_factor) {
+  PortableSymmetricQuantizeFloats(values, size, quantized_values, min_value,
+                                  max_value, scaling_factor);
+}
+
+void AsymmetricQuantizeFloats(const float* values, const int size,
+                              int8_t* quantized_values, float* scaling_factor,
+                              int32_t* offset) {
+  PortableAsymmetricQuantizeFloats(values, size, quantized_values,
+                                   scaling_factor, offset);
+}
+
+void MatrixBatchVectorMultiplyAccumulate(const float* matrix, int m_rows,
+                                         int m_cols, const float* vector,
+                                         int n_batch, float* result) {
+  PortableMatrixBatchVectorMultiplyAccumulate(matrix, m_rows, m_cols, vector,
+                                              n_batch, result);
+}
+
+void MatrixBatchVectorMultiplyAccumulate(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,
+                                         float* __restrict__ result) {
+  PortableMatrixBatchVectorMultiplyAccumulate(matrix, m_rows, m_cols, vector,
+                                              scaling_factors, n_batch, result);
+}
+
+void MatrixBatchVectorMultiplyAccumulate(
+    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) {
+  PortableMatrixBatchVectorMultiplyAccumulate(
+      matrix, m_rows, m_cols, vectors, scaling_factors, n_batch, result,
+      per_channel_scale, input_offset, scratch, row_sums, compute_row_sums,
+      context);
+}
+
+void MatrixBatchVectorMultiplyAccumulate(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) {
+  PortableMatrixBatchVectorMultiplyAccumulate(matrix, m_rows, m_cols, vector,
+                                              scaling_factors, n_batch, result);
+}
+
+void SparseMatrixBatchVectorMultiplyAccumulate1x4(
+    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) {
+  PortableSparseMatrixBatchVectorMultiplyAccumulate1x4(
+      matrix, segments, indices, m_rows, m_cols, vector, n_batch, result);
+}
+
+void SparseMatrixBatchVectorMultiplyAccumulate(
+    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) {
+  PortableSparseMatrixBatchVectorMultiplyAccumulate(
+      matrix, ledger, m_rows, m_cols, vector, n_batch, result);
+}
+
+void SparseMatrixBatchVectorMultiplyAccumulate1x16(
+    const int8_t* __restrict__ matrix, const int32_t* __restrict__ segments,
+    const int32_t* __restrict__ indices, int m_rows, int m_cols,
+    const int8_t* __restrict__ vector, const int32_t* __restrict__ bias_vector,
+    int n_batch, const int32_t input_offset, const int32_t output_multiplier,
+    const int32_t output_shift, const int32_t* per_channel_scale,
+    const int32_t* per_channel_shift, const int32_t output_offset,
+    const int32_t output_activation_min, const int32_t output_activation_max,
+
+    int8_t* __restrict__ result) {
+  PortableSparseMatrixBatchVectorMultiplyAccumulate1x16(
+      matrix, segments, indices, m_rows, m_cols, vector, bias_vector, n_batch,
+      input_offset, output_multiplier, output_shift, per_channel_scale,
+      per_channel_shift, output_offset, output_activation_min,
+      output_activation_max, result);
+}
+
+void SparseMatrixBatchVectorMultiplyAccumulate(
+    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,
+    const float* per_channel_scale) {
+  PortableSparseMatrixBatchVectorMultiplyAccumulate(
+      matrix, ledger, m_rows, m_cols, vectors, scaling_factors, n_batch, result,
+      per_channel_scale);
+}
+
+void MatrixBatchVectorMultiplyAccumulate(
+    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) {
+  PortableMatrixBatchVectorMultiplyAccumulate(
+      input, bias, input_to_gate_weights, multiplier, shift, n_batch, n_input,
+      n_output, output_zp, scratch, output, context);
+}
+
+void MatrixBatchVectorMultiplyAccumulate(
+    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) {
+  PortableMatrixBatchVectorMultiplyAccumulate(
+      input, bias, input_to_gate_weights, multiplier, shift, n_batch, n_input,
+      n_output, output_zp, scratch, output, context);
+}
+
+void MatrixScalarMultiplyAccumulate(const int8_t* matrix, int32_t scalar,
+                                    int32_t n_row, int32_t n_col,
+                                    int32_t* output) {
+  PortableMatrixScalarMultiplyAccumulate(matrix, scalar, n_row, n_col, output);
+}
+
+void MatrixBatchVectorMultiply(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) {
+  PortableMatrixBatchVectorMultiply(
+      input, input_zeropoint, input_to_gate_weights,
+      input_to_gate_effective_scale_a, input_to_gate_effective_scale_b, n_batch,
+      n_input, n_cell, gate_output, gate_output_zp);
+}
+
+void MatrixBatchVectorMultiply(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) {
+  PortableMatrixBatchVectorMultiply(hidden, hidden_to_output_weights,
+                                    proj_effective_scale_a,
+                                    proj_effective_scale_b, gate_bias, n_batch,
+                                    n_hidden, n_output, output_zp, proj_output);
+}
+
+void ApplyLayerNorm(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) {
+  PortableApplyLayerNorm(input, layer_norm_weights, bias, layer_norm_scale_a,
+                         layer_norm_scale_b, variance_limit, n_batch, n_input,
+                         output);
+}
+
+void ApplyLayerNormFloat(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) {
+  PortableApplyLayerNormFloat(input, layer_norm_weights, layer_norm_scale_a,
+                              layer_norm_scale_b, bias, n_batch, n_input,
+                              output);
+}
+
+void ApplySigmoid(const int16_t* input, int32_t n_batch, int32_t n_input,
+                  int16_t* output) {
+  PortableApplySigmoid(input, n_batch, n_input, output);
+}
+
+void ApplySigmoidFloat(const int16_t* input, int32_t n_batch, int32_t n_input,
+                       int16_t* output) {
+  PortableApplySigmoidFloat(input, n_batch, n_input, output);
+}
+
+void ApplyTanh(int32_t integer_bits, const int16_t* input, int32_t n_batch,
+               int32_t n_input, int16_t* output) {
+  PortableApplyTanh(integer_bits, input, n_batch, n_input, output);
+}
+
+void ApplyTanhFloat(const int16_t* input, int32_t n_batch, int32_t n_input,
+                    int32_t integer_bits, int16_t* output) {
+  PortableApplyTanhFloat(input, n_batch, n_input, integer_bits, output);
+}
+
+void CwiseMul(const int16_t* input_1, const int16_t* input_2, int n_batch,
+              int n_input, int shift, int16_t* output) {
+  PortableCwiseMul(input_1, input_2, n_batch, n_input, shift, output);
+}
+
+void CwiseMul(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) {
+  PortableCwiseMul(input_1, input_2, multiplier, shift, n_batch, n_input,
+                   output_zp, output);
+}
+
+void CwiseAdd(const int16_t* input_1, const int16_t* input_2, int n_batch,
+              int n_input, int16_t* output) {
+  PortableCwiseAdd(input_1, input_2, n_batch, n_input, output);
+}
+
+void CwiseClipping(float* vector, const int v_size,
+                   const float clipping_value) {
+  PortableCwiseClipping(vector, v_size, clipping_value);
+}
+
+void CwiseClipping(int16_t* vector, const int v_size,
+                   const int16_t clipping_value) {
+  PortableCwiseClipping(vector, v_size, clipping_value);
+}
+
+void CwiseClipping(int8_t* vector, const int v_size,
+                   const int8_t clipping_value) {
+  PortableCwiseClipping(vector, v_size, clipping_value);
+}
+
+void VectorBatchVectorCwiseProductAccumulate(const int16_t* vector, int v_size,
+                                             const int16_t* batch_vector,
+                                             int n_batch, int32_t multiplier,
+                                             int shift, int16_t* result) {
+  PortableVectorBatchVectorCwiseProductAccumulate(
+      vector, v_size, batch_vector, n_batch, multiplier, shift, result);
+}
+
+float VectorVectorDotProduct(const float* vector1, const float* vector2,
+                             int v_size) {
+  return PortableVectorVectorDotProduct(vector1, vector2, v_size);
+}
+
+void BatchVectorBatchVectorDotProduct(const int16_t* vector1,
+                                      const int16_t* vector2, int v_size,
+                                      int n_batch, int32_t* result) {
+  PortableBatchVectorBatchVectorDotProduct(vector1, vector2, v_size, n_batch,
+                                           result);
+}
+
+void Sub1Vector(const float* vector, int v_size, float* result) {
+  PortableSub1Vector(vector, v_size, result);
+}
+
+void Sub1Vector(const int16_t* vector, int v_size, int16_t* result) {
+  PortableSub1Vector(vector, v_size, result);
+}
+
+// Multiply all elements of vector with a scalar.
+void VectorScalarMultiply(const int8_t* vector, int v_size, float scale,
+                          float* result) {
+  PortableVectorScalarMultiply(vector, v_size, scale, result);
+}
+
+void ReductionSumVector(const float* input_vector, float* output_vector,
+                        int output_size, int reduction_size) {
+  PortableReductionSumVector(input_vector, output_vector, output_size,
+                             reduction_size);
+}
+
+void ReductionSumVector(const int32_t* input_vector, int32_t* output_vector,
+                        int output_size, int reduction_size) {
+  PortableReductionSumVector(input_vector, output_vector, output_size,
+                             reduction_size);
+}
+
+void ReductionSumVector(const int8_t* input_vector, int32_t* output_vector,
+                        int output_size, int reduction_size) {
+  PortableReductionSumVector(input_vector, output_vector, output_size,
+                             reduction_size);
+}
+
+void MeanStddevNormalization(const float* input_vector, float* output_vector,
+                             int v_size, int n_batch) {
+  PortableMeanStddevNormalization(input_vector, output_vector, v_size, n_batch);
+}
+
+void TwoGateSaturatingAdd(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) {
+  PortableTwoGateSaturatingAdd(
+      input, input_zp, recurrent, recurrent_zp, input_effective_scale_a,
+      input_effective_scale_b, recurrent_effective_scale_a,
+      recurrent_effective_scale_b, n_batch, n_cell, output);
+}
+
+}  // namespace tensor_utils
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PORTABLE_TENSOR_UTILS_H_

tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h

@@ -0,0 +1,248 @@
+/* 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;
+}
+
+// LINT.IfChange(portable_symmetric_quantize_floats)
+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);
+// LINT.ThenChange(//tensorflow/compiler/mlir/lite/quantization/lite/toco_legacy/portable_tensor_utils.h:portable_symmetric_quantize_floats)
+
+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 PortableSparseMatrixBatchVectorMultiplyAccumulate1x16(
+    const int8_t* __restrict__ matrix, const int32_t* __restrict__ segments,
+    const int32_t* __restrict__ indices, int m_rows, int m_cols,
+    const int8_t* __restrict__ vector, const int32_t* __restrict__ bias_vector,
+    int n_batch, const int32_t input_offset, const int32_t output_multiplier,
+    int32_t output_shift, const int32_t* per_channel_scale,
+    const int32_t* per_channel_shift, int32_t output_offset,
+    const int32_t output_activation_min, const int32_t output_activation_max,
+    int8_t* __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,
+    const float* per_channel_scale);
+
+// 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_

tensorflow/lite/kernels/internal/reference/prelu.h

@@ -0,0 +1,95 @@
+/* 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 <algorithm>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.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, typename U>
+inline void BroadcastPrelu4DSlow(
+    const PreluParams& params, const RuntimeShape& input_shape,
+    const T* input_data, const RuntimeShape& alpha_shape, const U* 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();
+
+  auto op = [&params, quantized_min, quantized_max](T input_val, U alpha_val) {
+    const int32_t input_value = params.input_offset + input_val;
+    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_val;
+      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));
+    return static_cast<T>(clamped_output);
+  };
+
+  BroadcastBinaryOpSimple(input_shape, input_data, alpha_shape, alpha_data,
+                          output_shape, output_data, op);
+}
+
+template <typename T, typename U>
+inline void Prelu(const PreluParams& params, const RuntimeShape& input_shape,
+                  const T* input_data, const RuntimeShape& alpha_shape,
+                  const U* 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_

tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h

@@ -0,0 +1,140 @@
+/* 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 <algorithm>
+
+#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_

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_

tensorflow/lite/kernels/internal/reference/reduce.h

@@ -0,0 +1,491 @@
+/* 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 <algorithm>
+
+#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;
+}
+
+inline void Mean(const tflite::MeanParams& op_params,
+                 const RuntimeShape& unextended_input_shape,
+                 const float* input_data,
+                 const RuntimeShape& unextended_output_shape,
+                 float* 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);
+    }
+  }
+}
+
+// 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,
+                               const int* input_dims, const int input_num_dims,
+                               T* output_data, int32_t output_multiplier,
+                               int output_shift, int32_t output_zero_point,
+                               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 int32_t kMinValue = std::numeric_limits<T>::min();
+  const int32_t kMaxValue = std::numeric_limits<T>::max();
+  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.
+  int64_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 > static_cast<size_t>(std::numeric_limits<int64_t>::max() /
+                                      num_elements_in_axis)) {
+      return false;
+    }
+    num_elements_in_axis *= current;
+  }
+
+  if (num_elements_in_axis == 0) {
+    return true;
+  }
+
+  // Readapt output rescaling when calculating the mean to integrate a
+  // 1/num_elements_in_axis multiplier.
+  if (!compute_sum) {
+    TFLITE_DCHECK_GE(num_elements_in_axis, 0);
+    int shift =
+        63 - CountLeadingZeros(static_cast<uint64_t>(num_elements_in_axis));
+    // To avoid any overflow risk 'shift' should be <= 32 and to satisfy
+    // 'MultiplyByQuantizedMultiplier' pre-conditions 'output_shift - shift'
+    // should be >= -31. Clamp the value at the price of some precision loss.
+    shift = std::min(shift, 32);
+    shift = std::min(shift, 31 + output_shift);
+    output_multiplier = static_cast<int32_t>(
+        (static_cast<int64_t>(output_multiplier) << shift) /
+        num_elements_in_axis);
+    output_shift = output_shift - shift;
+  }
+
+  for (size_t idx = 0; idx < num_outputs; ++idx) {
+    const U shifted_sum =
+        static_cast<U>(temp_sum[idx] - input_zero_point * num_elements_in_axis);
+    int32_t output = MultiplyByQuantizedMultiplier(
+                         shifted_sum, output_multiplier, output_shift) +
+                     output_zero_point;
+    output = std::min(std::max(output, kMinValue), kMaxValue);
+    output_data[idx] = static_cast<T>(output);
+  }
+  return true;
+}
+
+template <typename T, typename U>
+inline bool QuantizedMeanOrSumExtraArgs(
+    const T* input_data, int32_t input_zero_point, float input_scale,
+    const int* input_dims, const int input_num_dims, T* output_data,
+    float output_scale, int32_t output_multiplier, int output_shift,
+    int32_t output_zero_point, 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) {
+  return QuantizedMeanOrSum<T, U>(
+      input_data, input_zero_point, input_dims, input_num_dims, output_data,
+      output_multiplier, output_shift, output_zero_point, output_dims,
+      output_num_dims, axis, num_axis_dimensions, keep_dims, temp_index,
+      resolved_axis, temp_sum, compute_sum);
+}
+
+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_

tensorflow/lite/kernels/internal/reference/requantize.h

@@ -0,0 +1,70 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REQUANTIZE_H_
+
+#include <algorithm>
+
+#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_

tensorflow/lite/kernels/internal/reference/resize_bilinear.h

@@ -0,0 +1,233 @@
+/* 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;
+#if TFLITE_SINGLE_ROUNDING
+          const int64_t round = 1 << 19;
+          const T interpolation = static_cast<T>((output_20 + round) >> 20);
+#else
+          const int64_t round = (output_20 > 0) ? (1 << 19) : -(1 << 19);
+          const T interpolation =
+              static_cast<T>((output_20 + round) / (1 << 20));
+#endif  // TFLITE_SINGLE_ROUNDING
+          output_data[Offset(output_shape, b, y, x, c)] = interpolation;
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_RESIZE_BILINEAR_H_

tensorflow/lite/kernels/internal/reference/resize_nearest_neighbor.h

@@ -0,0 +1,102 @@
+/* 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 <algorithm>
+#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_

tensorflow/lite/kernels/internal/reference/reverse.h

@@ -0,0 +1,80 @@
+/* Copyright 2025 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_REVERSE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REVERSE_H_
+
+#include <algorithm>
+#include <array>
+#include <cstdint>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/runtime_shape.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename Scalar>
+void Reverse(std::array<int32_t, 8>& axes, int num_axes,
+             const RuntimeShape& input_shape, const Scalar* input_data,
+             Scalar* output_data) {
+  ruy::profiler::ScopeLabel label("Reverse");
+  bool is_upper = (axes[num_axes - 1] == input_shape.DimensionsCount() - 1);
+  bool is_lower = (axes[0] == 0);
+  int rank = input_shape.DimensionsCount();
+  if (is_upper && is_lower) {
+    std::reverse_copy(input_data, input_data + input_shape.FlatSize(),
+                      output_data);
+    return;
+  } else {
+    int32_t min_dim = axes[0];
+    int32_t max_dim = axes[num_axes - 1];
+    int upper_size = 1;
+    for (int i = 0; i < min_dim; ++i) {
+      upper_size *= input_shape.Dims(i);
+    }
+    int lower_size = 1;
+    for (int i = max_dim + 1; i < rank; ++i) {
+      lower_size *= input_shape.Dims(i);
+    }
+    int middle_size = 1;
+    for (int i = min_dim; i <= max_dim; ++i) {
+      middle_size *= input_shape.Dims(i);
+    }
+
+    if (lower_size > 1) {
+      for (int i = 0; i < upper_size; ++i) {
+        for (int j = 0; j < middle_size; ++j) {
+          Scalar* src =
+              (Scalar*)input_data + (i * (middle_size) + j) * lower_size;
+          Scalar* dst =
+              (Scalar*)output_data +
+              (i * (middle_size) + (middle_size - j - 1)) * lower_size;
+          memcpy(dst, src, lower_size * sizeof(Scalar));
+        }
+      }
+    } else {
+      for (int i = 0; i < upper_size; ++i) {
+        std::reverse_copy(input_data + i * (middle_size),
+                          input_data + i * middle_size + middle_size,
+                          output_data + i * (middle_size));
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_REVERSE_H_

tensorflow/lite/kernels/internal/reference/round.h

@@ -0,0 +1,52 @@
+/* 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;
+  }
+}
+
+template <typename Scalar>
+inline void Round(const RuntimeShape& input_shape, const Scalar* input_data,
+                  const RuntimeShape& output_shape, Scalar* 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] = static_cast<Scalar>(RoundToNearest(input_data[i]));
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_

tensorflow/lite/kernels/internal/reference/select.h

@@ -0,0 +1,253 @@
+/* Copyright 2022 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_SELECT_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SELECT_H_
+
+#include <algorithm>
+#include <cmath>
+#include <cstring>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename D, typename T>
+void Select(const RuntimeShape& input_condition_shape,
+            const D* input_condition_data, const RuntimeShape& input_x_shape,
+            const T* input_x_data, const RuntimeShape& input_y_shape,
+            const T* input_y_data, const RuntimeShape& output_shape,
+            T* output_data) {
+  ruy::profiler::ScopeLabel label("Select");
+  int64_t flatsize;
+  // Allow select operator executions on mixed scalar tensors and one element
+  // tensors.
+  if (input_condition_shape.FlatSize() == 1 && input_x_shape.FlatSize() == 1 &&
+      input_y_shape.FlatSize() == 1 && output_shape.FlatSize() == 1) {
+    flatsize = 1;
+  } else {
+    flatsize = MatchingFlatSize(input_condition_shape, input_x_shape,
+                                input_y_shape, output_shape);
+  }
+  for (int64_t i = 0; i < flatsize; ++i) {
+    output_data[i] =
+        input_condition_data[i] ? input_x_data[i] : input_y_data[i];
+  }
+}
+
+template <typename D, typename T>
+void RankOneSelect(const RuntimeShape& input_condition_shape,
+                   const D* input_condition_data,
+                   const RuntimeShape& input_x_shape, const T* input_x_data,
+                   const RuntimeShape& input_y_shape, const T* input_y_data,
+                   const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("Select/RankOneSelect");
+  const int64_t outer_size = input_condition_shape.FlatSize();
+  int64_t inner_size;
+  if (input_condition_shape.DimensionsCount() == 0) {
+    inner_size = MatchingFlatSize(input_x_shape, input_y_shape, output_shape);
+  } else {
+    TFLITE_DCHECK_EQ(
+        MatchingDim(input_x_shape, 0, input_y_shape, 0, output_shape, 0),
+        outer_size);
+    inner_size =
+        MatchingFlatSizeSkipDim(input_x_shape, 0, input_y_shape, output_shape);
+  }
+
+  int64_t offset = 0;
+  for (int64_t i = 0; i < outer_size; i++) {
+    const T* input_data = input_condition_data[i] ? input_x_data : input_y_data;
+    memcpy(output_data + offset, input_data + offset, inner_size * sizeof(T));
+    offset += inner_size;
+  }
+}
+
+template <typename D, typename T>
+void RunSelectOp(const D* cond, const T* x, const T* y, T* output,
+                 const size_t* cond_stride, const size_t* x_stride,
+                 const size_t* y_stride, const size_t* output_stride,
+                 const size_t* output_shape, int dim) {
+  TFLITE_DCHECK_GE(dim, 0);
+  size_t output_shape_0 = output_shape[dim];
+  size_t output_stride_0 = output_stride[dim];
+  size_t cond_stride_0 = cond_stride[dim];
+  size_t x_stride_0 = x_stride[dim];
+  size_t y_stride_0 = y_stride[dim];
+  if (dim == 0) {
+    TFLITE_DCHECK_EQ(output_stride_0, 1);
+    if (cond_stride_0 == 0) {
+      if (*cond) {
+        if (x_stride_0 == 0) {
+          std::fill_n(output, output_shape_0, *x);
+        } else {
+          TFLITE_DCHECK_EQ(x_stride_0, 1);
+          std::memcpy(output, x, output_shape_0 * sizeof(T));
+        }
+      } else {
+        if (y_stride_0 == 0) {
+          std::fill_n(output, output_shape_0, *y);
+        } else {
+          TFLITE_DCHECK_EQ(y_stride_0, 1);
+          std::memcpy(output, y, output_shape_0 * sizeof(T));
+        }
+      }
+    } else {
+      TFLITE_DCHECK_EQ(cond_stride_0, 1);
+      if (x_stride_0 == 0 && y_stride_0 == 0) {
+        for (size_t i = 0; i < output_shape_0; ++i) {
+          output[i] = cond[i] ? *x : *y;
+        }
+      } else if (x_stride_0 == 0) {
+        TFLITE_DCHECK_EQ(y_stride_0, 1);
+        for (size_t i = 0; i < output_shape_0; ++i) {
+          output[i] = cond[i] ? *x : y[i];
+        }
+      } else if (y_stride_0 == 0) {
+        TFLITE_DCHECK_EQ(x_stride_0, 1);
+        for (size_t i = 0; i < output_shape_0; ++i) {
+          output[i] = cond[i] ? x[i] : *y;
+        }
+      } else {
+        TFLITE_DCHECK_EQ(x_stride_0, 1);
+        TFLITE_DCHECK_EQ(y_stride_0, 1);
+        for (size_t i = 0; i < output_shape_0; ++i) {
+          output[i] = cond[i] ? x[i] : y[i];
+        }
+      }
+    }
+  } else {
+    dim -= 1;
+    for (size_t i = 0; i < output_shape_0; ++i) {
+      RunSelectOp(cond, x, y, output, cond_stride, x_stride, y_stride,
+                  output_stride, output_shape, dim);
+      cond += cond_stride_0;
+      x += x_stride_0;
+      y += y_stride_0;
+      output += output_stride_0;
+    }
+  }
+}
+
+template <typename D, typename T>
+inline void BroadcastSelectSimple(const RuntimeShape& cond_shape,
+                                  const D* cond_data,
+                                  const RuntimeShape& x_shape, const T* x_data,
+                                  const RuntimeShape& y_shape, const T* y_data,
+                                  const RuntimeShape& output_shape,
+                                  T* output_data) {
+  constexpr int kMaxRank = 8;
+  const int dims_count = std::max(
+      output_shape.DimensionsCount(),
+      std::max(cond_shape.DimensionsCount(),
+               std::max(x_shape.DimensionsCount(), y_shape.DimensionsCount())));
+  if (dims_count <= 0) {
+    *output_data = *cond_data ? *x_data : *y_data;
+    return;
+  }
+
+  TFLITE_DCHECK_LE(dims_count, kMaxRank);
+
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(dims_count, output_shape);
+  const RuntimeShape extended_cond_shape =
+      RuntimeShape::ExtendedShape(dims_count, cond_shape);
+  const RuntimeShape extended_x_shape =
+      RuntimeShape::ExtendedShape(dims_count, x_shape);
+  const RuntimeShape extended_y_shape =
+      RuntimeShape::ExtendedShape(dims_count, y_shape);
+
+  size_t cond_strides[kMaxRank];
+  size_t x_strides[kMaxRank];
+  size_t y_strides[kMaxRank];
+  size_t o_strides[kMaxRank];
+  size_t o_shape[kMaxRank];
+
+  size_t cond_accum_stride = 1;
+  size_t x_accum_stride = 1;
+  size_t y_accum_stride = 1;
+  size_t o_accum_stride = 1;
+  int next_dim_idx = -1;
+  for (int i = dims_count - 1; i >= 0; --i) {
+    const int cond_dim = extended_cond_shape.Dims(i);
+    const int x_dim = extended_x_shape.Dims(i);
+    const int y_dim = extended_y_shape.Dims(i);
+    const int output_dim = extended_output_shape.Dims(i);
+    if (cond_dim <= 0 || x_dim <= 0 || y_dim <= 0 || output_dim <= 0) {
+      // Empty operation.
+      return;
+    }
+    size_t cond_stride =
+        (cond_dim == 1 && output_dim != 1) ? 0 : cond_accum_stride;
+    size_t x_stride = (x_dim == 1 && output_dim != 1) ? 0 : x_accum_stride;
+    size_t y_stride = (y_dim == 1 && output_dim != 1) ? 0 : y_accum_stride;
+    size_t o_stride = o_accum_stride;
+
+    if (next_dim_idx >= 0 &&
+        CanFuseLoops(output_dim, cond_dim, cond_stride, cond_accum_stride,
+                     cond_strides[next_dim_idx]) &&
+        CanFuseLoops(output_dim, x_dim, x_stride, x_accum_stride,
+                     x_strides[next_dim_idx]) &&
+        CanFuseLoops(output_dim, y_dim, y_stride, y_accum_stride,
+                     y_strides[next_dim_idx]) &&
+        CanFuseLoops(output_dim, output_dim, o_stride, o_accum_stride,
+                     o_strides[next_dim_idx])) {
+      // This dimension can be fused into one loop with the previous
+      // dimension.
+      o_shape[next_dim_idx] *= output_dim;
+    } else {
+      ++next_dim_idx;
+      cond_strides[next_dim_idx] = cond_stride;
+      x_strides[next_dim_idx] = x_stride;
+      y_strides[next_dim_idx] = y_stride;
+      o_strides[next_dim_idx] = o_stride;
+      o_shape[next_dim_idx] = output_dim;
+    }
+
+    cond_accum_stride *= cond_dim;
+    x_accum_stride *= x_dim;
+    y_accum_stride *= y_dim;
+    o_accum_stride *= output_dim;
+  }
+
+  RunSelectOp(cond_data, x_data, y_data, output_data, cond_strides, x_strides,
+              y_strides, o_strides, o_shape, next_dim_idx);
+}
+
+template <typename D, typename T>
+void BroadcastSelect5DSlow(const RuntimeShape& input_condition_shape,
+                           const D* input_condition_data,
+                           const RuntimeShape& input_x_shape,
+                           const T* input_x_data,
+                           const RuntimeShape& input_y_shape,
+                           const T* input_y_data,
+                           const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("Select/BroadcastSelectSlow");
+  TFLITE_DCHECK_LE(input_condition_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_LE(input_x_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_LE(input_y_shape.DimensionsCount(), 5);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), 5);
+
+  BroadcastSelectSimple(input_condition_shape, input_condition_data,
+                        input_x_shape, input_x_data, input_y_shape,
+                        input_y_data, output_shape, output_data);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SELECT_H_

tensorflow/lite/kernels/internal/reference/slice.h

@@ -0,0 +1,176 @@
+/* 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 <cstdint>
+#include <vector>
+
+#include "tensorflow/lite/core/c/common.h"
+#include "tensorflow/lite/kernels/internal/portable_tensor.h"
+#include "tensorflow/lite/kernels/internal/portable_tensor_utils.h"
+#include "tensorflow/lite/kernels/internal/runtime_shape.h"
+#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void Slice(const std::vector<int>& begins,
+                  const RuntimeShape& input_shape,
+                  const RuntimeShape& output_shape,
+                  SequentialTensorWriter<T>* writer) {
+  const int dims = input_shape.DimensionsCount();
+  std::vector<int> input_strides(dims);
+  std::vector<int> output_strides(dims);
+  if (dims == 0) {
+    writer->Write(0);
+    return;
+  }
+  input_strides[dims - 1] = 1;
+  output_strides[dims - 1] = 1;
+  for (int i = dims - 2; i >= 0; --i) {
+    input_strides[i] = input_strides[i + 1] * input_shape.Dims(i + 1);
+    output_strides[i] = output_strides[i + 1] * output_shape.Dims(i + 1);
+  }
+  for (int output_index = 0; output_index < output_shape.FlatSize();
+       ++output_index) {
+    int remaining_index = output_index;
+    int input_index = 0;
+    for (int dim = 0; dim < dims; ++dim) {
+      const int coordinate = remaining_index / output_strides[dim];
+      remaining_index %= output_strides[dim];
+      input_index += (begins[dim] + coordinate) * input_strides[dim];
+    }
+    writer->Write(input_index);
+  }
+}
+
+template <typename T>
+inline void Slice(const std::vector<int>& begins,
+                  const RuntimeShape& input_shape, const T* input_data,
+                  const RuntimeShape& output_shape, T* output_data) {
+  SequentialTensorWriter<T> writer(input_data, output_data);
+  return Slice(begins, input_shape, output_shape, &writer);
+}
+
+template <typename T>
+inline void Slice(const std::vector<int>& begins,
+                  const RuntimeShape& input_shape, const TfLiteTensor* input,
+                  const RuntimeShape& output_shape, TfLiteTensor* output) {
+  SequentialTensorWriter<T> writer(input, output);
+  return Slice(begins, input_shape, output_shape, &writer);
+}
+
+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);
+}
+
+inline void SliceInt4(const tflite::SliceParams& op_params,
+                      const RuntimeShape& input_shape,
+                      const TfLiteTensor* input,
+                      const RuntimeShape& output_shape, TfLiteTensor* output) {
+  const int num_input_elements = input_shape.FlatSize();
+  std::vector<int8_t> unpacked_input(num_input_elements);
+  tensor_utils::UnpackPackedIntToInt8(GetTensorData<int8_t>(input),
+                                      num_input_elements, 4,
+                                      unpacked_input.data());
+
+  const int num_output_elements = output_shape.FlatSize();
+  std::vector<int8_t> unpacked_output(num_output_elements);
+
+  reference_ops::Slice<int8_t>(op_params, input_shape, unpacked_input.data(),
+                               output_shape, unpacked_output.data());
+
+  tensor_utils::PackInt8IntoDenseInt(unpacked_output.data(),
+                                     num_output_elements, 4,
+                                     GetTensorData<int8_t>(output));
+}
+
+inline void SliceInt4(const std::vector<int>& begins,
+                      const RuntimeShape& input_shape,
+                      const TfLiteTensor* input,
+                      const RuntimeShape& output_shape, TfLiteTensor* output) {
+  const int num_input_elements = input_shape.FlatSize();
+  std::vector<int8_t> unpacked_input(num_input_elements);
+  tensor_utils::UnpackPackedIntToInt8(GetTensorData<int8_t>(input),
+                                      num_input_elements, 4,
+                                      unpacked_input.data());
+
+  const int num_output_elements = output_shape.FlatSize();
+  std::vector<int8_t> unpacked_output(num_output_elements);
+
+  reference_ops::Slice<int8_t>(begins, input_shape, unpacked_input.data(),
+                               output_shape, unpacked_output.data());
+
+  tensor_utils::PackInt8IntoDenseInt(unpacked_output.data(),
+                                     num_output_elements, 4,
+                                     GetTensorData<int8_t>(output));
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SLICE_H_

tensorflow/lite/kernels/internal/reference/softmax.h

@@ -0,0 +1,236 @@
+/* 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 <algorithm>
+#include <limits>
+#include <type_traits>
+
+#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 {
+
+template <typename T,
+          typename std::enable_if<!std::is_integral<T>::value, int>::type = 0>
+inline void Softmax(const SoftmaxParams& params,
+                    const RuntimeShape& input_shape, const T* input_data,
+                    const RuntimeShape& output_shape, 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) {
+    T max = std::numeric_limits<T>::lowest();
+    for (int c = 0; c < depth; ++c) {
+      max = std::max(max, input_data[i * depth + c]);
+    }
+
+    float sum = 0.f;
+    for (int c = 0; c < depth; ++c) {
+      const float exp_c =
+          std::exp((static_cast<float>(input_data[i * depth + c]) -
+                    static_cast<float>(max)) *
+                   static_cast<float>(params.beta));
+      output_data[i * depth + c] = static_cast<T>(exp_c);
+      sum += exp_c;
+    }
+
+    for (int c = 0; c < depth; ++c) {
+      output_data[i * depth + c] =
+          static_cast<T>(static_cast<float>(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));
+
+    const int exponent = num_bits_over_unit + 31 - (sizeof(OutputT) * 8);
+    TFLITE_CHECK(0 <= exponent && exponent <= 31);
+
+    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(), exponent);
+
+        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 LUTLookup(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 =
+        LUTLookup(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_

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_

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_

tensorflow/lite/kernels/internal/reference/sparse_ops/fully_connected.h

@@ -0,0 +1,46 @@
+/* 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_SPARSE_OPS_FULLY_CONNECTED_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SPARSE_OPS_FULLY_CONNECTED_H_
+
+#include "tensorflow/lite/kernels/internal/reference/fully_connected.h"
+#include "tensorflow/lite/kernels/internal/utils/sparsity_format_converter.h"
+
+namespace tflite {
+namespace reference_ops {
+
+// Convert weights to dense format and run dense fully connected.
+inline void FullyConnectedSparseWeight(
+    const TfLiteSparsity& sparsity, 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) {
+  std::vector<int> weights_shape_vector(weights_shape.DimensionsCount());
+  for (int i = 0; i < weights_shape.DimensionsCount(); i++) {
+    weights_shape_vector[i] = weights_shape.Dims(i);
+  }
+  tflite::internal::sparsity::FormatConverter<float> converter(
+      weights_shape_vector, sparsity);
+  converter.SparseToDense(weights_data);
+  const std::vector<float>& dense_weights_data = converter.GetData();
+  FullyConnected(params, input_shape, input_data, weights_shape,
+                 dense_weights_data.data(), bias_shape, bias_data, output_shape,
+                 output_data);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SPARSE_OPS_FULLY_CONNECTED_H_

tensorflow/lite/kernels/internal/reference/strided_slice.h

@@ -0,0 +1,267 @@
+/* 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 <functional>
+#include <vector>
+
+#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 {
+
+struct DynamicStridedSliceParams {
+  std::vector<int32_t> start_indices;
+  std::vector<int32_t> stop_indices;
+  std::vector<int32_t> strides;
+  uint32_t begin_mask = 0;
+  uint32_t end_mask = 0;
+  uint32_t shrink_axis_mask = 0;
+  bool offset = false;
+};
+
+inline bool AxisMask(uint32_t mask, int axis) {
+  return mask & (uint32_t{1} << axis);
+}
+
+inline int StartForAxis(const DynamicStridedSliceParams& params,
+                        const RuntimeShape& input_shape, int axis) {
+  const int axis_size = input_shape.Dims(axis);
+  int start = params.start_indices[axis];
+  const int stride = params.strides[axis];
+  if (start < 0) {
+    start += axis_size;
+  }
+  if (stride > 0) {
+    start = strided_slice::Clamp(start, 0, axis_size);
+  } else {
+    start = strided_slice::Clamp(start, -1, axis_size - 1);
+  }
+  if (AxisMask(params.begin_mask, axis)) {
+    start = stride > 0 ? 0 : axis_size - 1;
+  }
+  return start;
+}
+
+inline int EndForAxis(const DynamicStridedSliceParams& params,
+                      const RuntimeShape& input_shape, int axis, int start) {
+  const bool shrink_axis = AxisMask(params.shrink_axis_mask, axis);
+  const int axis_size = input_shape.Dims(axis);
+  if (shrink_axis) {
+    return start >= axis_size ? start : start + 1;
+  }
+  int end = params.stop_indices[axis];
+  if (params.offset) {
+    end += start;
+  }
+  const int stride = params.strides[axis];
+  if (end < 0) {
+    end += axis_size;
+  }
+  if (stride > 0) {
+    end = strided_slice::Clamp(end, 0, axis_size);
+  } else {
+    end = strided_slice::Clamp(end, -1, axis_size - 1);
+  }
+  if (AxisMask(params.end_mask, axis)) {
+    end = stride > 0 ? axis_size : -1;
+  }
+  return end;
+}
+
+template <typename T>
+inline void StridedSlice(const DynamicStridedSliceParams& op_params,
+                         const RuntimeShape& input_shape,
+                         const RuntimeShape& output_shape,
+                         SequentialTensorWriter<T>* writer) {
+  ruy::profiler::ScopeLabel label("StridedSlice");
+  const int dims = input_shape.DimensionsCount();
+  std::vector<int> starts(dims);
+  std::vector<int> stops(dims);
+  std::vector<int> input_strides(dims);
+  if (dims == 0) {
+    writer->Write(0);
+    return;
+  }
+  input_strides[dims - 1] = 1;
+  for (int i = dims - 2; i >= 0; --i) {
+    input_strides[i] = input_strides[i + 1] * input_shape.Dims(i + 1);
+  }
+  for (int axis = 0; axis < dims; ++axis) {
+    starts[axis] = StartForAxis(op_params, input_shape, axis);
+    stops[axis] = EndForAxis(op_params, input_shape, axis, starts[axis]);
+  }
+
+  auto loop_condition = [](int index, int stop, int stride) {
+    return stride > 0 ? index < stop : index > stop;
+  };
+  std::function<void(int, int)> write_slice = [&](int axis, int input_index) {
+    if (axis == dims) {
+      writer->Write(input_index);
+      return;
+    }
+    for (int offset = starts[axis];
+         loop_condition(offset, stops[axis], op_params.strides[axis]);
+         offset += op_params.strides[axis]) {
+      write_slice(axis + 1, input_index + offset * input_strides[axis]);
+    }
+  };
+  write_slice(/*axis=*/0, /*input_index=*/0);
+}
+
+template <typename T>
+inline void StridedSlice(const DynamicStridedSliceParams& op_params,
+                         const RuntimeShape& input_shape, const T* input_data,
+                         const RuntimeShape& output_shape, T* output_data) {
+  SequentialTensorWriter<T> writer(input_data, output_data);
+  StridedSlice<T>(op_params, input_shape, output_shape, &writer);
+}
+
+template <typename T>
+inline void StridedSlice(const DynamicStridedSliceParams& op_params,
+                         const RuntimeShape& input_shape,
+                         const TfLiteTensor* input,
+                         const RuntimeShape& output_shape,
+                         TfLiteTensor* output) {
+  SequentialTensorWriter<T> writer(input, output);
+  StridedSlice<T>(op_params, input_shape, output_shape, &writer);
+}
+
+template <typename T>
+inline void StridedSlice(const tflite::StridedSliceParams& op_params,
+                         const RuntimeShape& unextended_input_shape,
+                         const RuntimeShape& unextended_output_shape,
+                         SequentialTensorWriter<T>* writer) {
+  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 =
+      strided_slice::StridedSliceStartForAxis(params_copy, input_shape, 0);
+  const int stop_0 = strided_slice::StridedSliceEndForAxis(
+      params_copy, input_shape, 0, start_0);
+  const int start_1 =
+      strided_slice::StridedSliceStartForAxis(params_copy, input_shape, 1);
+  const int stop_1 = strided_slice::StridedSliceEndForAxis(
+      params_copy, input_shape, 1, start_1);
+  const int start_2 =
+      strided_slice::StridedSliceStartForAxis(params_copy, input_shape, 2);
+  const int stop_2 = strided_slice::StridedSliceEndForAxis(
+      params_copy, input_shape, 2, start_2);
+  const int start_3 =
+      strided_slice::StridedSliceStartForAxis(params_copy, input_shape, 3);
+  const int stop_3 = strided_slice::StridedSliceEndForAxis(
+      params_copy, input_shape, 3, start_3);
+  const int start_4 =
+      strided_slice::StridedSliceStartForAxis(params_copy, input_shape, 4);
+  const int stop_4 = strided_slice::StridedSliceEndForAxis(
+      params_copy, input_shape, 4, start_4);
+
+  auto lc = [&](int end, int stride, int index) {
+    if (stride < 0) {
+      return index > end;
+    } else {
+      return index < end;
+    }
+  };
+  // With a static_cast it is not possible to initialize
+  // a variable of type 'const int *'
+  // with an rvalue of type 'const int32_t *' (aka 'const long *').
+  // reinterpret_cast is required to handle this casting.
+  const int* shape = reinterpret_cast<const int*>(input_shape.DimsData());
+  const int* stride = reinterpret_cast<const int*>(params_copy.strides);
+  const bool inner_stride_is_1 = params_copy.strides[4] == 1;
+
+  for (int offset_0 = start_0; lc(stop_0, stride[0], offset_0);
+       offset_0 += stride[0]) {
+    for (int offset_1 = start_1; lc(stop_1, stride[1], offset_1);
+         offset_1 += stride[1]) {
+      for (int offset_2 = start_2; lc(stop_2, stride[2], offset_2);
+           offset_2 += stride[2]) {
+        for (int offset_3 = start_3; lc(stop_3, stride[3], offset_3);
+             offset_3 += stride[3]) {
+          // When the stride is 1, the inner loop is equivalent to the
+          // optimized slice inner loop. Otherwise, it is identical to the
+          // strided_slice reference implementation inner loop.
+          if (inner_stride_is_1) {
+            const int len = stop_4 - start_4;
+            int index = start_4 + offset_3 * shape[4] +
+                        offset_2 * shape[3] * shape[4] +
+                        offset_1 * shape[2] * shape[3] * shape[4] +
+                        offset_0 * shape[1] * shape[2] * shape[3] * shape[4];
+            if (len > 0) {
+              writer->WriteN(index, len);
+            }
+          } else {
+            for (int offset_4 = start_4; lc(stop_4, stride[4], offset_4);
+                 offset_4 += stride[4]) {
+              int index = offset_4 + offset_3 * shape[4] +
+                          offset_2 * shape[3] * shape[4] +
+                          offset_1 * shape[2] * shape[3] * shape[4] +
+                          offset_0 * shape[1] * shape[2] * shape[3] * shape[4];
+              writer->Write(index);
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+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_

tensorflow/lite/kernels/internal/reference/string_comparisons.h

@@ -0,0 +1,85 @@
+/* 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_STRING_COMPARISONS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRING_COMPARISONS_H_
+
+#include "tensorflow/lite/core/c/common.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/reference/broadcast_loop.h"
+#include "tensorflow/lite/kernels/internal/reference/comparisons.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/string_util.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline bool StringRefEqualFn(const StringRef& lhs, const StringRef& rhs) {
+  if (lhs.len != rhs.len) return false;
+  for (int i = 0; i < lhs.len; ++i) {
+    if (lhs.str[i] != rhs.str[i]) return false;
+  }
+  return true;
+}
+
+inline bool StringRefNotEqualFn(const StringRef& lhs, const StringRef& rhs) {
+  return !StringRefEqualFn(lhs, rhs);
+}
+
+inline void ComparisonStringImpl(bool (*F)(const StringRef&, const StringRef&),
+                                 const RuntimeShape& input1_shape,
+                                 const TfLiteTensor* input1,
+                                 const RuntimeShape& input2_shape,
+                                 const TfLiteTensor* input2,
+                                 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) {
+    const auto lhs = GetString(input1, i);
+    const auto rhs = GetString(input2, i);
+    output_data[i] = F(lhs, rhs);
+  }
+}
+
+struct TfLiteTensorStringPointer {
+  const TfLiteTensor* tensor;
+  int index;
+
+  TfLiteTensorStringPointer operator+(int offset) const {
+    return {tensor, index + offset};
+  }
+  StringRef operator*() const { return GetString(tensor, index); }
+  StringRef operator[](int offset) const {
+    return GetString(tensor, index + offset);
+  }
+};
+
+inline void BroadcastComparison4DSlowStringImpl(
+    bool (*F)(const StringRef&, const StringRef&),
+    const RuntimeShape& unextended_input1_shape, const TfLiteTensor* input1,
+    const RuntimeShape& unextended_input2_shape, const TfLiteTensor* input2,
+    const RuntimeShape& unextended_output_shape, bool* output_data) {
+  auto op = [F](const StringRef& a, const StringRef& b) { return F(a, b); };
+  BroadcastBinaryOpSimple(
+      unextended_input1_shape, TfLiteTensorStringPointer{input1, 0},
+      unextended_input2_shape, TfLiteTensorStringPointer{input2, 0},
+      unextended_output_shape, output_data, op);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_STRING_COMPARISONS_H_

tensorflow/lite/kernels/internal/reference/sub.h

@@ -0,0 +1,257 @@
+/* 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 "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/reference/broadcast_loop.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T, typename F>
+inline void BroadcastSubCommon(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,
+                               F binary_func) {
+  auto op = [&params, binary_func](T a, T b) {
+    return binary_func(a, b, params);
+  };
+  BroadcastBinaryOpSimple(input1_shape, input1_data, input2_shape, input2_data,
+                          output_shape, output_data, op);
+}
+
+template <typename T>
+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/T");
+  BroadcastSubCommon<T>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](T input1_val, T input2_val, const ArithmeticParams& params) {
+        T activation_min, activation_max;
+        GetActivationParams(params, &activation_min, &activation_max);
+        return ActivationFunctionWithMinMax(input1_val - input2_val,
+                                            activation_min, activation_max);
+      });
+}
+
+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");
+  BroadcastSubCommon<int16_t>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](int16_t input1_val, int16_t input2_val,
+         const ArithmeticParams& params) {
+        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));
+        return static_cast<int16_t>(clamped_output);
+      });
+}
+
+template <typename T>
+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");
+  BroadcastSubCommon<T>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](T input1_val, T input2_val, const ArithmeticParams& params) {
+        const int32_t shifted_input1_val =
+            (params.input1_offset + input1_val) * (1 << params.left_shift);
+        const int32_t shifted_input2_val =
+            (params.input2_offset + 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));
+        return static_cast<T>(clamped_output);
+      });
+}
+
+// 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) {
+  BroadcastSubCommon<T>(
+      params, input1_shape, input1_data, input2_shape, input2_data,
+      output_shape, output_data,
+      [](T input1_val, T input2_val, const ArithmeticParams& params) {
+        return input1_val - input2_val;
+      });
+}
+
+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_

tensorflow/lite/kernels/internal/reference/svdf.h

@@ -0,0 +1,249 @@
+/* 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_SVDF_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SVDF_H_
+
+#include <stdint.h>
+
+#include <algorithm>
+#include <limits>
+
+#include "tensorflow/lite/core/c/builtin_op_data.h"
+#include "tensorflow/lite/core/c/common.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
+#include "tensorflow/lite/kernels/internal/tensor_utils.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+// SVDF op that compresses a fully connected op via low-rank matrix
+// factorization. See https://research.google.com/pubs/archive/43813.pdf for
+// details.
+
+namespace tflite {
+namespace reference_ops {
+
+static inline void ApplyTimeWeightsBiasAndActivation(
+    int batch_size, int memory_size, int num_filters, int num_units, int rank,
+    const float* const __restrict__ weights_time_data,
+    const float* const __restrict__ bias_ptr, TfLiteFusedActivation activation,
+    float* const __restrict__ state_ptr, float* const __restrict__ scratch_ptr,
+    float* const __restrict__ output_ptr) {
+  // Compute matmul(state, weights_time).
+  for (int b = 0; b < batch_size; ++b) {
+    float* state_ptr_batch = state_ptr + b * memory_size * num_filters;
+    float* scratch_ptr_batch = scratch_ptr + b * num_filters;
+    tensor_utils::BatchVectorBatchVectorDotProduct(
+        weights_time_data, state_ptr_batch, memory_size, num_filters,
+        scratch_ptr_batch);
+  }
+
+  // Reduction sum.
+  tensor_utils::ReductionSumVector(scratch_ptr, output_ptr,
+                                   batch_size * num_units, rank);
+  // Add bias if provided.
+  if (bias_ptr) {
+    tensor_utils::VectorBatchVectorAdd(bias_ptr, num_units, batch_size,
+                                       output_ptr);
+  }
+
+  // Apply activation.
+  tensor_utils::ApplyActivationToVector(output_ptr, batch_size * num_units,
+                                        activation, output_ptr);
+}
+
+inline void EvalIntegerSVDF(
+    const TfLiteSVDFParams* params, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& weights_feature_shape,
+    const int8_t* weights_feature_data, const RuntimeShape& weights_time_shape,
+    const int16_t* weights_time_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, int16_t* state_data,
+    const RuntimeShape& output_shape, int8_t* output_data,
+    int32_t* scratch_data, int32_t* output_temp_data, int32_t scale_1_a,
+    int scale_1_b, int32_t scale_2_a, int scale_2_b, int32_t input_zp,
+    int32_t output_zp) {
+  const int n_rank = params->rank;
+  const int n_batch = input_shape.Dims(0);
+  const int n_input = input_shape.Dims(1);
+  const int n_filter = weights_feature_shape.Dims(0);
+  const int n_unit = n_filter / n_rank;
+  const int n_memory = weights_time_shape.Dims(1);
+
+  // Left shift the activation_state.
+  // std::copy is fine for overlapping ranges if the output is outside of the
+  // input range. (This is not true for copy_n.)
+  std::copy(state_data + 1, state_data + n_batch * n_memory * n_filter,
+            state_data);
+
+  // Feature matmul.
+  // Note: no need to clear the latest activation, matmul is not accumulative.
+  {
+    const int32_t output_max = std::numeric_limits<int16_t>::max();
+    const int32_t output_min = std::numeric_limits<int16_t>::min();
+    int16_t* result_in_batch = state_data + (n_memory - 1);
+    for (int b = 0; b < n_batch; b++) {
+      const int8_t* matrix_data = weights_feature_data;
+      for (int r = 0; r < n_filter; r++) {
+        int32_t dot_prod = 0;
+        const int8_t* vector_in_batch = input_data + b * n_input;
+        for (int c = 0; c < n_input; c++) {
+          dot_prod += *matrix_data++ * (*vector_in_batch++ - input_zp);
+        }
+        dot_prod =
+            MultiplyByQuantizedMultiplier(dot_prod, scale_1_a, scale_1_b);
+        dot_prod = std::min(std::max(output_min, dot_prod), output_max);
+        // This assumes state is symmetrically quantized. Otherwise last bit of
+        // state should be initialized to its zero point and accumulate the
+        // dot_prod.
+        // Equivalent as the following:
+        //     result_in_batch = zero point, which happens to be zero.
+        //     result_in_batch += dot_prod.
+        *result_in_batch = dot_prod;
+        result_in_batch += n_memory;
+      }
+    }
+  }
+
+  // Time.
+  {
+    for (int b = 0; b < n_batch; ++b) {
+      const int16_t* state_data_batch = state_data + b * n_memory * n_filter;
+      int32_t* scratch_data_batch = scratch_data + b * n_filter;
+      tensor_utils::BatchVectorBatchVectorDotProduct(
+          weights_time_data, state_data_batch, n_memory, n_filter,
+          scratch_data_batch);
+    }
+  }
+
+  // Reduce, add bias, rescale, activation.
+  {
+    // Reduce.
+    tensor_utils::ReductionSumVector(scratch_data, output_temp_data,
+                                     n_batch * n_unit, n_rank);
+    // Add bias.
+    if (bias_data) {
+      tensor_utils::VectorBatchVectorAdd(bias_data, n_unit, n_batch,
+                                         output_temp_data);
+    }
+    // Rescale.
+    const int32_t output_max = std::numeric_limits<int8_t>::max();
+    const int32_t output_min = std::numeric_limits<int8_t>::min();
+    for (int i = 0; i < n_batch * n_unit; ++i) {
+      int32_t x1 = output_temp_data[i];
+      int32_t x2 = MultiplyByQuantizedMultiplier(x1, scale_2_a, scale_2_b);
+      int32_t x3 = x2 + output_zp;
+      int32_t x4 = std::min(std::max(output_min, x3), output_max);
+      output_data[i] = static_cast<int8_t>(x4);
+    }
+  }
+}
+
+inline void EvalFloatSVDF(
+    const TfLiteSVDFParams* params, const RuntimeShape& input_shape,
+    const float* input_data, const RuntimeShape& weights_feature_shape,
+    const float* weights_feature_data, const RuntimeShape& weights_time_shape,
+    const float* weights_time_data, const RuntimeShape& bias_shape,
+    const float* bias_data, float* scratch_data, float* state_data,
+    const RuntimeShape& output_shape, float* output_data) {
+  const int rank = params->rank;
+  const int batch_size = input_shape.Dims(0);
+  const int input_size = input_shape.Dims(1);
+  const int num_filters = weights_feature_shape.Dims(0);
+  const int num_units = num_filters / rank;
+  const int memory_size = weights_time_shape.Dims(1);
+
+  // Left shift the activation_state.
+  // std::copy is fine for overlapping ranges if the output is outside of the
+  // input range. (This is not true for copy_n.)
+  std::copy(state_data + 1, state_data + batch_size * memory_size * num_filters,
+            state_data);
+
+  // Clear scratch (the matmul is accumulative).
+  std::fill_n(scratch_data, batch_size * num_filters, 0.0f);
+
+  // Compute conv1d(inputs, weights_feature).
+  tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+      weights_feature_data, num_filters, input_size, input_data, batch_size,
+      scratch_data);
+
+  // Copy the latest activation from scratch into activation_state:
+  // The last, i.e. (memory_size-1)th entry for each batch, and filter.
+  for (int i = 0; i < batch_size * num_filters; ++i) {
+    state_data[i * memory_size + memory_size - 1] = scratch_data[i];
+  }
+
+  ApplyTimeWeightsBiasAndActivation(
+      batch_size, memory_size, num_filters, num_units, rank, weights_time_data,
+      bias_data, params->activation, state_data, scratch_data, output_data);
+}
+
+inline void EvalHybridSVDF(
+    const TfLiteSVDFParams* params, const RuntimeShape& input_shape,
+    const float* input_data, const RuntimeShape& weights_feature_shape,
+    const int8_t* weights_feature_data, const float weights_feature_scale,
+    const RuntimeShape& weights_time_shape, const float* weights_time_data,
+    const RuntimeShape& bias_shape, const float* bias_data, float* scratch,
+    float* scaling_factors, int8_t* quantized_input, float* state,
+    const RuntimeShape& output_shape, float* output_data, int32_t* zero_points,
+    int32_t* row_sums, bool* compute_row_sums) {
+  const int rank = params->rank;
+  const int batch_size = input_shape.Dims(0);
+  const int input_size = input_shape.Dims(1);
+  const int num_filters = weights_feature_shape.Dims(0);
+  const int num_units = num_filters / rank;
+  const int memory_size = weights_time_shape.Dims(1);
+
+  // Left shift the activation_state.
+  // std::copy is fine for overlapping ranges if the output is outside of the
+  // input range. (This is not true for copy_n.)
+  std::copy(state + 1, state + batch_size * memory_size * num_filters, state);
+
+  // Clear scratch (the matmul is accumulative).
+  std::fill_n(scratch, batch_size * num_filters, 0.0f);
+
+  if (!tensor_utils::IsZeroVector(input_data, batch_size * input_size)) {
+    // Quantize input from float to int8_t.
+    tensor_utils::BatchQuantizeFloats(
+        input_data, batch_size, input_size, quantized_input, scaling_factors,
+        zero_points, params->asymmetric_quantize_inputs);
+    for (int b = 0; b < batch_size; ++b) {
+      scaling_factors[b] *= weights_feature_scale;
+    }
+
+    // Compute conv1d(inputs, weights_feature).
+    tensor_utils::MatrixBatchVectorMultiplyAccumulate(
+        weights_feature_data, num_filters, input_size, quantized_input,
+        scaling_factors, batch_size, scratch,
+        /*per_channel_scale=*/nullptr, zero_points,
+        reinterpret_cast<int32_t*>(scratch), row_sums, compute_row_sums,
+        /*context=*/nullptr);
+  }
+  // Copy the latest activation from scratch into activation_state:
+  // The last, i.e. (memory_size-1)th entry for each batch, and filter.
+  for (int i = 0; i < batch_size * num_filters; ++i) {
+    state[i * memory_size + memory_size - 1] = scratch[i];
+  }
+
+  // TODO(b/174275776): can optimize hybrid case ~5% by unrolling loop in
+  // applying time weights so that the inner loop multiplies eight elements at
+  // a time.
+  ApplyTimeWeightsBiasAndActivation(
+      batch_size, memory_size, num_filters, num_units, rank, weights_time_data,
+      bias_data, params->activation, state, scratch, output_data);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SVDF_H_

tensorflow/lite/kernels/internal/reference/tanh.h

@@ -0,0 +1,128 @@
+/* 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 {
+
+template <typename T>
+inline void Tanh(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] = static_cast<T>(std::tanh(input_data[i]));
+  }
+}
+
+// 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_

tensorflow/lite/kernels/internal/reference/transpose.h

@@ -0,0 +1,206 @@
+/* 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 <array>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+namespace transpose_internal {
+
+// Recursively explores all the dimensions of the output tensor and writes the
+// corresponding input tensor data.
+//
+// - depth: the current depth of the recursion.
+// - dims: tensor dimension count, also `perm` size.
+// - perm: permutation array.
+// - input_data: Running input data pointer. If depth == num_dims-1, this points
+//               to the first element of the last dimension to traverse.
+// - input_stride: Reverse partial product of input shapes.
+// - output_data: Running output data pointer. If depth == num_dims-1, this
+//                points to the first element of the last dimension to traverse.
+// - output_stride: Reverse partial product of output shapes.
+// - output_shape: Shape of the output tensor.
+//
+// ## Algorithm explanation
+//
+// Assume a 3D tensor T with a shape of [I, J, K] stored in row major order.
+// T[i, j, k] is at position `i*J*K + j*K + k` in the tensor buffer.
+//
+// If we want to go through the whole tensor iteratively, we can use loops.
+//
+// ```
+// for(i = 0; i < I; ++i) {
+//   for(j = 0; j < J; ++j) {
+//     for(k = 0; k < K; ++k) {
+//        T.data[i*J*K + j*K + k] = ...
+//     }
+//   }
+// }
+// ```
+//
+// We can also compute the offset as we go through the loops.
+//
+// ```
+// stride_i = K * J;
+// stride_j = K;
+// stride_k = 1;
+// for(i = 0; i < I; ++i) {
+//   offset_i = i * stride_i;
+//   offset_j = 0;
+//   for(j = 0; j < J; ++j) {
+//     offset_j += stride_j;
+//     offset_k = 0;
+//     for(k = 0; k < K; ++k) {
+//        offset_k += stride_k;
+//        T.data[offset_i + offset_j + offset_k] = ...
+//     }
+//   }
+// }
+// ```
+//
+// This nicely extends to a recursive version which is the base of this
+// algorithm and supports any number of dimensions.
+//
+// ```
+// shape = [I, J, K]
+// strides = [K*J, K, 1]
+// void recurse(T* data, shape, strides, depth = 0) {
+//   if(depth == shape.size) {
+//     *data = ...
+//   } else {
+//     for(a = 0; a < shape[depth]; ++a) {
+//       recurse(data, shape, strides, depth+1);
+//       data += strides[depth];
+//     }
+//   }
+// }
+// ```
+template <typename T>
+void TransposeImpl(const int depth, const int dims, const int32_t* perm,
+                   const T* input_data, const int* input_stride, T* output_data,
+                   const int* output_stride, const int32_t* output_shape) {
+  const int dimension_size = output_shape[depth];
+  if (depth == dims - 1) {
+    const int loop_stride = input_stride[perm[depth]];
+    for (int i = 0; i < dimension_size; ++i) {
+      output_data[i] = *input_data;
+      input_data += loop_stride;
+    }
+  } else {
+    for (int i = 0; i < dimension_size; ++i) {
+      TransposeImpl(depth + 1, dims, perm, input_data, input_stride,
+                    output_data, output_stride, output_shape);
+
+      input_data += input_stride[perm[depth]];
+      output_data += output_stride[depth];
+    }
+  }
+}
+
+// Compile-time switch to get the storage type of the transposition.
+template <int Size>
+struct TransposeStorageType;
+
+template <>
+struct TransposeStorageType<1> {
+  using type = int8_t;
+};
+
+template <>
+struct TransposeStorageType<2> {
+  using type = int16_t;
+};
+
+template <>
+struct TransposeStorageType<4> {
+  using type = int32_t;
+};
+
+template <>
+struct TransposeStorageType<8> {
+  using type = int64_t;
+};
+
+// Sets up the stride arrays for the recursive transpose algorithm.
+//
+// Implementation notes:
+//
+// This is a reverse partial product. We could use standard algorithms to
+// implement this but the result is not a readable and is tricky to get right
+// because the first element must be set to 1, which leads to offset
+// shenanigans:
+//
+// ```
+//   stride[dims - 1] = 1;
+//   std::partial_sum(std::make_reverse_iterator(shape + dims),
+//                    std::make_reverse_iterator(shape + 1),
+//                    stride.rend() - input_rank + 1, std::multiplies());
+// ```
+//
+// Note that Abseil isn't used in kernels implementation. That would make the
+// above solution more readable.
+inline void SetupTransposeStrides(
+    std::array<int, kTransposeMaxDimensions>& stride, const int32_t* shape,
+    const int dims) {
+  stride[dims - 1] = 1;
+  for (int i = dims - 2; i >= 0; --i) {
+    stride[i] = stride[i + 1] * shape[i + 1];
+  }
+}
+
+}  // namespace transpose_internal
+
+// Copies a tensor to an other buffer and permutes its dimensions.
+//
+// Note: template parameter N is not used anymore. It is kept for API
+// compatibility with TFLite micro.
+template <typename T, int N = kTransposeMaxDimensions>
+void Transpose(const TransposeParams& params, const RuntimeShape& input_shape,
+               const T* input_data, const RuntimeShape& output_shape,
+               T* output_data) {
+  if (input_shape.FlatSize() == 0) {
+    return;
+  }
+  using transpose_internal::SetupTransposeStrides;
+  using transpose_internal::TransposeImpl;
+  using transpose_internal::TransposeStorageType;
+  // 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.
+  using StorageType = typename TransposeStorageType<sizeof(T)>::type;
+  const StorageType* const input_data_storage =
+      reinterpret_cast<const StorageType*>(input_data);
+  StorageType* const output_data_storage =
+      reinterpret_cast<StorageType*>(output_data);
+
+  const int dims = input_shape.DimensionsCount();
+  std::array<int, kTransposeMaxDimensions> input_stride, output_stride;
+  SetupTransposeStrides(input_stride, input_shape.DimsData(), dims);
+  SetupTransposeStrides(output_stride, output_shape.DimsData(), dims);
+  TransposeImpl(0, dims, &params.perm[0], input_data_storage,
+                input_stride.data(), output_data_storage, output_stride.data(),
+                output_shape.DimsData());
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_H_

tensorflow/lite/kernels/internal/reference/transpose_conv.h

@@ -0,0 +1,322 @@
+/* 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 <algorithm>
+
+#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);
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  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;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  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) {
+          float acc = output_data[Offset(output_shape, batch, out_y, out_x,
+                                         out_channel)];
+          if (bias_data) acc += bias_data[out_channel];
+
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              ActivationFunctionWithMinMax(acc, output_activation_min,
+                                           output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+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);
+        }
+      }
+    }
+  }
+}
+
+inline void HybridTransposeConv(
+    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 float* per_channel_scale, int32_t* input_offset) {
+  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);
+
+  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 float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  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) {
+    const float scaling_factor = scaling_factors_ptr[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)) {
+                  int32_t input_value = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_value =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  int32_t acc =
+                      (input_value - input_offset[batch]) * filter_value;
+                  output_data[Offset(output_shape, batch, out_y, out_x,
+                                     out_channel)] +=
+                      acc * per_channel_scale[out_channel] * scaling_factor;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  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) {
+          float acc = output_data[Offset(output_shape, batch, out_y, out_x,
+                                         out_channel)];
+          if (bias_data) acc += bias_data[out_channel];
+
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              ActivationFunctionWithMinMax(acc, output_activation_min,
+                                           output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_TRANSPOSE_CONV_H_