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

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

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

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

@@ -0,0 +1,51 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite_micro {
+
+namespace reference_ops {
+
+inline float RoundToNearest(float value) {
+  auto floor_val = std::floor(value);
+  auto diff = value - floor_val;
+  if ((diff < 0.5f) ||
+      ((diff == 0.5f) && (static_cast<int>(floor_val) % 2 == 0))) {
+    return floor_val;
+  } else {
+    return floor_val = floor_val + 1.0f;
+  }
+}
+
+inline void Round(const RuntimeShape& input_shape, const float* input_data,
+                  const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    // Note that this implementation matches that of tensorFlow tf.round
+    // and corresponds to the bankers rounding method.
+    // cfenv (for fesetround) is not yet supported universally on Android, so
+    // using a work around.
+    output_data[i] = RoundToNearest(input_data[i]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite_micro
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ROUND_H_

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

@@ -0,0 +1,151 @@
+/* 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 <cmath>
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite_micro {
+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 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);
+
+  NdArrayDesc<5> desc_condition;
+  NdArrayDesc<5> desc_x;
+  NdArrayDesc<5> desc_y;
+  NdArrayDesc<5> desc_output;
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(5, output_shape);
+  CopyDimsToDesc(extended_output_shape, &desc_output);
+  NdArrayDescsForElementwiseBroadcast(input_condition_shape, input_x_shape,
+                                      input_y_shape, &desc_condition, &desc_x,
+                                      &desc_y);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest
+  // stride, typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for
+  // the best cache behavior.
+  for (int n = 0; n < desc_output.extents[0]; ++n) {
+    int out_idx_n = desc_output.extents[1] * n;
+    int cond_idx_n = desc_condition.strides[0] * n;
+    int in_idx1_n = desc_x.strides[0] * n;
+    int in_idx2_n = desc_y.strides[0] * n;
+    for (int b = 0; b < desc_output.extents[1]; ++b) {
+      int out_idx_b = (out_idx_n + b) * desc_output.extents[2];
+      int cond_idx_b = cond_idx_n + desc_condition.strides[1] * b;
+      int in_idx1_b = in_idx1_n + desc_x.strides[1] * b;
+      int in_idx2_b = in_idx2_n + desc_y.strides[1] * b;
+      for (int y = 0; y < desc_output.extents[2]; ++y) {
+        int out_idx_y = (out_idx_b + y) * desc_output.extents[3];
+        int cond_idx_y = cond_idx_b + desc_condition.strides[2] * y;
+        int in_idx1_y = in_idx1_b + desc_x.strides[2] * y;
+        int in_idx2_y = in_idx2_b + desc_y.strides[2] * y;
+        for (int x = 0; x < desc_output.extents[3]; ++x) {
+          int out_idx = (out_idx_y + x) * desc_output.extents[4];
+          int cond_idx = cond_idx_y + desc_condition.strides[3] * x;
+          int in_idx1 = in_idx1_y + desc_x.strides[3] * x;
+          int in_idx2 = in_idx2_y + desc_y.strides[3] * x;
+          for (int c = 0; c < desc_output.extents[4]; ++c) {
+            output_data[out_idx] = input_condition_data[cond_idx]
+                                       ? input_x_data[in_idx1]
+                                       : input_y_data[in_idx2];
+            out_idx++;
+            cond_idx += desc_condition.strides[4];
+            in_idx1 += desc_x.strides[4];
+            in_idx2 += desc_y.strides[4];
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SELECT_H_

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

@@ -0,0 +1,80 @@
+/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SLICE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SLICE_H_
+
+#include "tensorflow/lite/kernels/internal/portable_tensor.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite_micro {
+
+namespace reference_ops {
+
+template <typename T>
+inline void Slice(const tflite_micro::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_micro::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_micro::SliceParams& op_params,
+                  const RuntimeShape& input_shape, const TfLiteTensor* input,
+                  const RuntimeShape& output_shape, TfLiteTensor* output) {
+  SequentialTensorWriter<T> writer(input, output);
+  return Slice(op_params, input_shape, output_shape, &writer);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite_micro
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_SLICE_H_