xmos-ai-tools 1.3.2.dev80__py3-none-macosx_10_15_universal2.whl

xmos_ai_tools/runtime/include/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_micro {
+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_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_

xmos_ai_tools/runtime/include/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_micro {
+
+namespace reference_ops {
+
+inline void L2Normalization(const tflite_micro::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_micro::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_micro
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_

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

@@ -0,0 +1,69 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LEAKY_RELU_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LEAKY_RELU_H_
+
+#include <algorithm>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite_micro {
+namespace reference_ops {
+
+inline void LeakyRelu(const tflite_micro::LeakyReluParams& params,
+                      const RuntimeShape& input_shape, const float* input_data,
+                      const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    const float val = input_data[i];
+    // Note that alpha might be > 1 or < 0, so we don't use std::max here.
+    output_data[i] = val > 0 ? val : val * params.alpha;
+  }
+}
+
+template <typename T>
+inline void QuantizeLeakyRelu(const LeakyReluParams& params,
+                              const RuntimeShape& input_shape,
+                              const T* input_data,
+                              const RuntimeShape& output_shape,
+                              T* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  static const int32_t quantized_min = std::numeric_limits<T>::min();
+  static const int32_t quantized_max = std::numeric_limits<T>::max();
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input_value = input_data[i] - params.input_offset;
+    int32_t unclamped_output;
+    if (input_value >= 0) {
+      unclamped_output = params.output_offset +
+                         MultiplyByQuantizedMultiplier(
+                             input_value, params.output_multiplier_identity,
+                             params.output_shift_identity);
+    } else {
+      unclamped_output = params.output_offset +
+                         MultiplyByQuantizedMultiplier(
+                             input_value, params.output_multiplier_alpha,
+                             params.output_shift_alpha);
+    }
+    const T clamped_output =
+        std::min(quantized_max, std::max(quantized_min, unclamped_output));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LEAKY_RELU_H_

xmos_ai_tools/runtime/include/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_micro {
+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_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOG_SOFTMAX_H_