xmos-ai-tools 1.3.2.dev180__py3-none-macosx_10_15_universal2.whl

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h

@@ -0,0 +1,244 @@
+/* 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_micro {
+
+// Not all backends support CpuBackendContext usage, so forward declare to avoid
+// pulling in its implementation.
+class CpuBackendContext;
+
+namespace tensor_utils {
+
+template <typename T>
+bool PortableIsZeroVector(const T* vector, int v_size) {
+  for (int i = 0; i < v_size; ++i) {
+    if (vector[i] != 0) {
+      return false;
+    }
+  }
+  return true;
+}
+
+void PortableSymmetricQuantizeFloats(const float* values, const int size,
+                                     int8_t* quantized_values, float* min_value,
+                                     float* max_value, float* scaling_factor);
+
+void PortableSymmetricQuantizeFloats(const float* values, const int size,
+                                     int8_t* quantized_values, float min_value,
+                                     float max_value, float* scaling_factor);
+
+void PortableAsymmetricQuantizeFloats(const float* values, const int size,
+                                      int8_t* quantized_values,
+                                      float* scaling_factor, int32_t* offset);
+
+// Multiply a matrix by a batch vector, and store results in a batch-size
+// vector.
+void PortableMatrixBatchVectorMultiplyAccumulate(const float* matrix,
+                                                 int m_rows, int m_cols,
+                                                 const float* vector,
+                                                 int n_batch, float* result);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vectors, const float* scaling_factors,
+    int n_batch, float* __restrict__ result);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vectors, const float* scaling_factors,
+    int n_batch, float* __restrict__ result, const float* per_channel_scale,
+    const int32_t* input_offset, int32_t* scratch, int32_t* row_sums,
+    bool* compute_row_sums, CpuBackendContext* context);
+
+void PortableMatrixBatchVectorMultiplyAccumulate(
+    const int8_t* __restrict__ matrix, const int m_rows, const int m_cols,
+    const int8_t* __restrict__ vector, const float* scaling_factors,
+    int n_batch, int32_t* scratch, float* __restrict__ result,
+    CpuBackendContext* context);
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate1x4(
+    const float* __restrict__ matrix, const int32_t* __restrict__ segments,
+    const int32_t* __restrict__ indices, int m_rows, int m_cols,
+    const float* __restrict__ vector, int n_batch, float* __restrict__ result);
+
+void PortableSparseMatrixBatchVectorMultiplyAccumulate(
+    const float* __restrict__ matrix, const uint8_t* __restrict__ ledger,
+    int m_rows, int m_cols, const float* __restrict__ vector, int n_batch,
+    float* __restrict__ result);
+
+void 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 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);
+
+// 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_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PORTABLE_TENSOR_UTILS_IMPL_H_

xmos_ai_tools/runtime/include/tensorflow/lite/kernels/internal/reference/prelu.h

@@ -0,0 +1,111 @@
+/* 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/types.h"
+
+namespace tflite_micro {
+
+namespace reference_ops {
+
+// Broadcast prelu to output_shape for quantized uint8_t/int8_t data.
+template <typename T>
+inline void BroadcastPrelu4DSlow(
+    const PreluParams& params, const RuntimeShape& input_shape,
+    const T* input_data, const RuntimeShape& alpha_shape, const T* alpha_data,
+    const RuntimeShape& output_shape, T* output_data) {
+  TFLITE_DCHECK_LE(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(alpha_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), 4);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input_shape, alpha_shape, &desc1, &desc2);
+
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          int output_index = Offset(extended_output_shape, b, y, x, c);
+          int input_index = SubscriptToIndex(desc1, b, y, x, c);
+          const int32_t input_value =
+              params.input_offset + input_data[input_index];
+          int32_t output_value;
+          if (input_value >= 0) {
+            output_value = MultiplyByQuantizedMultiplier(
+                input_value, params.output_multiplier_1, params.output_shift_1);
+          } else {
+            auto alpha_index = SubscriptToIndex(desc2, b, y, x, c);
+            const int32_t alpha_value =
+                params.alpha_offset + alpha_data[alpha_index];
+
+            output_value = MultiplyByQuantizedMultiplier(
+                input_value * alpha_value, params.output_multiplier_2,
+                params.output_shift_2);
+          }
+          output_value += params.output_offset;
+
+          const int32_t quantized_min = std::numeric_limits<T>::min();
+          const int32_t quantized_max = std::numeric_limits<T>::max();
+          const int32_t clamped_output =
+              std::min(quantized_max, std::max(quantized_min, output_value));
+          output_data[output_index] = static_cast<T>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+inline void Prelu(const PreluParams& params, const RuntimeShape& input_shape,
+                  const T* input_data, const RuntimeShape& alpha_shape,
+                  const T* alpha_data, const RuntimeShape& output_shape,
+                  T* output_data) {
+  const int32_t quantized_min = std::numeric_limits<T>::min();
+  const int32_t quantized_max = std::numeric_limits<T>::max();
+
+  const int flat_size =
+      MatchingElementsSize(input_shape, alpha_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input_value = params.input_offset + input_data[i];
+    int32_t output_value;
+    if (input_value >= 0) {
+      output_value = MultiplyByQuantizedMultiplier(
+          input_value, params.output_multiplier_1, params.output_shift_1);
+    } else {
+      const int32_t alpha_value = params.alpha_offset + alpha_data[i];
+
+      output_value = MultiplyByQuantizedMultiplier(input_value * alpha_value,
+                                                   params.output_multiplier_2,
+                                                   params.output_shift_2);
+    }
+    output_value += params.output_offset;
+
+    const int32_t clamped_output =
+        std::min(quantized_max, std::max(quantized_min, output_value));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_

xmos_ai_tools/runtime/include/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_micro {
+
+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_micro::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_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_

xmos_ai_tools/runtime/include/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_micro {
+
+namespace reference_ops {
+
+template <typename InputT, typename OutputT>
+inline void AffineQuantize(const tflite_micro::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_micro::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_micro
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_