Sync from tflite-micro. (#164)

Author: TFLM-bot

src/tensorflow/lite/core/c/common.cpp

@@ -219,11 +219,11 @@ TfLiteStatus TfLiteTensorCopy(const TfLiteTensor* src, TfLiteTensor* dst) {
   return kTfLiteOk;
 }
 
-void TfLiteTensorResizeMaybeCopy(size_t num_bytes, TfLiteTensor* tensor,
-                                 bool preserve_data) {
+TfLiteStatus TfLiteTensorResizeMaybeCopy(size_t num_bytes, TfLiteTensor* tensor,
+                                         bool preserve_data) {
   if (tensor->allocation_type != kTfLiteDynamic &&
       tensor->allocation_type != kTfLitePersistentRo) {
-    return;
+    return kTfLiteOk;
   }
 #ifdef TF_LITE_TENSORFLOW_PROFILER
   tflite::PauseHeapMonitoring(/*pause=*/true);
@@ -258,9 +258,15 @@ void TfLiteTensorResizeMaybeCopy(size_t num_bytes, TfLiteTensor* tensor,
   tflite::PauseHeapMonitoring(/*pause=*/false);
 #endif
   tensor->bytes = num_bytes;
+  if (tensor->data.data == nullptr && num_bytes != 0) {
+    // We are done allocating but tensor is pointing to null and a valid size
+    // was requested, so we error.
+    return kTfLiteError;
+  }
+  return kTfLiteOk;
 }
 
-void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor) {
+TfLiteStatus TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor) {
   return TfLiteTensorResizeMaybeCopy(num_bytes, tensor, true);
 }
 #endif  // TF_LITE_STATIC_MEMORY

src/tensorflow/lite/core/c/common.h

@@ -42,6 +42,7 @@ limitations under the License.
 #ifndef TENSORFLOW_LITE_CORE_C_COMMON_H_
 #define TENSORFLOW_LITE_CORE_C_COMMON_H_
 
+#include <stdarg.h>
 #include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -648,23 +649,26 @@ void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims,
 TfLiteStatus TfLiteTensorCopy(const TfLiteTensor* src, TfLiteTensor* dst);
 
 // Change the size of the memory block owned by `tensor` to `num_bytes`.
-// Tensors with allocation types other than kTfLiteDynamic will be ignored.
+// Tensors with allocation types other than `kTfLiteDynamic` will be ignored and
+// a kTfLiteOk will be returned.
 // `tensor`'s internal data buffer will be assigned a pointer
 // which can safely be passed to free or realloc if `num_bytes` is zero.
-// Behaviour is undefined if `tensor` is NULL.
 // If `preserve_data` is true, tensor data will be unchanged in the range from
-// the start of the region up to the minimum of the old and new sizes.
-void TfLiteTensorResizeMaybeCopy(size_t num_bytes, TfLiteTensor* tensor,
-                                 bool preserve_data);
+// the start of the region up to the minimum of the old and new sizes. In the
+// case of NULL tensor, or an error allocating new memory, returns
+// `kTfLiteError`.
+TfLiteStatus TfLiteTensorResizeMaybeCopy(size_t num_bytes, TfLiteTensor* tensor,
+                                         bool preserve_data);
 
 // Change the size of the memory block owned by `tensor` to `num_bytes`.
-// Tensors with allocation types other than kTfLiteDynamic will be ignored.
+// Tensors with allocation types other than kTfLiteDynamic will be ignored and
+// a kTfLiteOk will be returned.
 // `tensor`'s internal data buffer will be assigned a pointer
 // which can safely be passed to free or realloc if `num_bytes` is zero.
-// Behaviour is undefined if `tensor` is NULL.
 // Tensor data will be unchanged in the range from the start of the region up to
-// the minimum of the old and new sizes.
-void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor);
+// the minimum of the old and new sizes. In the case
+// of NULL tensor, or an error allocating new memory, returns `kTfLiteError`.
+TfLiteStatus TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor);
 #endif  // TF_LITE_STATIC_MEMORY
 
 // WARNING: This is an experimental interface that is subject to change.

src/tensorflow/lite/kernels/internal/common.h

@@ -377,40 +377,49 @@ inline Integer FloorLog2(Integer n) {
   }
 }
 
-// The size of the LUT depends on the type of input. For uint8 and int8 inputs
-// we use a 256 entries LUT to map all the values in the (u)int8 range. For
-// int16 inputs the high 9 bits are used for indexing and the 7 remaining bits
-// are used for interpolation. We thus use a 513-entries LUT for int16 cases,
-// 512 for the 9-bit indexing and 1 extra entry to interpolate the last value.
-template <typename T>
-constexpr int LUTSize() {
-  static_assert(std::is_same<T, uint8_t>::value ||
-                    std::is_same<T, int8_t>::value ||
-                    std::is_same<T, int16_t>::value,
-                "Only LUTs with uint8, int8 or int16 inputs are supported.");
-  // As per c++11: constexpr methods cannot have more than one return statement.
-  return (std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value)
-             ? 256
-             : 513;
+namespace detail {
+
+// LUTPopulate takes an optional type-erased transform_params to allow passing
+// extra parameters to the transform function pointer. const void* is used
+// instead of std::function to be compatible with TFLite Micro
+template <typename FloatT, typename Func>
+inline typename std::enable_if<std::is_same<Func, FloatT (*)(FloatT)>::value,
+                               FloatT>::type
+LUTTransform(Func transform, const void* /*transform_params*/, FloatT value) {
+  static_assert(std::is_floating_point<FloatT>::value,
+                "FloatT must be a floating-point type.");
+  return transform(value);
+}
+
+template <typename FloatT, typename Func>
+inline typename std::enable_if<
+    std::is_same<Func, FloatT (*)(FloatT, const void*)>::value, FloatT>::type
+LUTTransform(Func transform, const void* transform_params, FloatT value) {
+  static_assert(std::is_floating_point<FloatT>::value,
+                "FloatT must be a floating-point type.");
+  return transform(value, transform_params);
 }
 
 // Use the same LUT generation code for both uint8_t and int8_t. Int8_t indexes
 // will be directly casted to uint8_t, the int8 LUT will thus be ordered as [0,
 // 1, ..., 127, -128, ..., -2, -1] instead of [-128, -127, ..., -1, 0, 1, ...,
 // 126, 127].
-template <typename T>
-inline typename std::enable_if<std::is_same<T, uint8_t>::value ||
-                                   std::is_same<T, int8_t>::value,
-                               void>::type
-LUTPopulate(float input_scale, int32_t input_zero_point, float output_scale,
-            int32_t output_zero_point, float (*transform)(float), T* lut) {
+template <typename T, typename Func>
+inline void LUTPopulateInt8(float input_scale, int32_t input_zero_point,
+                            float output_scale, int32_t output_zero_point,
+                            Func transform, const void* transform_params,
+                            T* lut) {
+  static_assert(
+      std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value,
+      "T must be an uint8 or int8 type.");
   uint8_t* lut_uint8 = reinterpret_cast<uint8_t*>(lut);
   const float inverse_scale = 1 / output_scale;
   int32_t maxval = std::numeric_limits<T>::max();
   int32_t minval = std::numeric_limits<T>::min();
   for (int32_t val = minval; val <= maxval; ++val) {
     const float dequantized = input_scale * (val - input_zero_point);
-    const float transformed = transform(dequantized);
+    const float transformed =
+        LUTTransform(transform, transform_params, dequantized);
     const float rescaled = TfLiteRound(transformed * inverse_scale);
     const int32_t quantized =
         static_cast<int32_t>(rescaled + output_zero_point);
@@ -421,10 +430,11 @@ LUTPopulate(float input_scale, int32_t input_zero_point, float output_scale,
 
 // Keep floating-point type configurable for backward compatibility. float
 // should be used for FloatT by default.
-template <typename T, typename FloatT>
-inline typename std::enable_if<std::is_same<T, int16_t>::value, void>::type
-LUTPopulate(FloatT input_scale, int32_t input_zero_point, FloatT output_scale,
-            int32_t output_zero_point, FloatT (*transform)(FloatT), T* lut) {
+template <typename FloatT, typename Func>
+inline void LUTPopulateInt16(FloatT input_scale, int32_t input_zero_point,
+                             FloatT output_scale, int32_t output_zero_point,
+                             Func transform, const void* transform_params,
+                             int16_t* lut) {
   static_assert(std::is_floating_point<FloatT>::value,
                 "FloatT must be a floating-point type.");
   const FloatT input_min =
@@ -440,16 +450,21 @@ LUTPopulate(FloatT input_scale, int32_t input_zero_point, FloatT output_scale,
   const FloatT step = (input_max - input_min) / nb_steps;
   const FloatT half_step = step / 2;
   const FloatT output_scaling_inv =
-      static_cast<FloatT>(std::numeric_limits<T>::max() -
-                          std::numeric_limits<T>::min() + 1) /
+      static_cast<FloatT>(std::numeric_limits<int16_t>::max() -
+                          std::numeric_limits<int16_t>::min() + 1) /
       (output_max - output_min);
-  const FloatT table_min = static_cast<FloatT>(std::numeric_limits<T>::min());
-  const FloatT table_max = static_cast<FloatT>(std::numeric_limits<T>::max());
+  const FloatT table_min =
+      static_cast<FloatT>(std::numeric_limits<int16_t>::min());
+  const FloatT table_max =
+      static_cast<FloatT>(std::numeric_limits<int16_t>::max());
 
   for (int i = 0; i < nb_steps; i++) {
-    const FloatT val = transform(input_min + i * step);
-    const FloatT val_midpoint = transform(input_min + i * step + half_step);
-    const FloatT val_next = transform(input_min + (i + 1) * step);
+    const FloatT val =
+        LUTTransform<FloatT>(transform, transform_params, input_min + i * step);
+    const FloatT val_midpoint = LUTTransform<FloatT>(
+        transform, transform_params, input_min + i * step + half_step);
+    const FloatT val_next = LUTTransform<FloatT>(transform, transform_params,
+                                                 input_min + (i + 1) * step);
 
     const FloatT sample_val = TfLiteRound(val * output_scaling_inv);
     const FloatT midpoint_interp_val =
@@ -460,54 +475,84 @@ LUTPopulate(FloatT input_scale, int32_t input_zero_point, FloatT output_scale,
     const FloatT midpoint_err = midpoint_interp_val - midpoint_val;
     const FloatT bias = TfLiteRound(midpoint_err / 2);
 
-    lut[i] = static_cast<T>(std::min<FloatT>(
+    lut[i] = static_cast<int16_t>(std::min<FloatT>(
         std::max<FloatT>(sample_val - bias, table_min), table_max));
   }
 
-  lut[nb_steps] = static_cast<T>(std::min<FloatT>(
-      std::max<FloatT>(TfLiteRound(transform(input_max) * output_scaling_inv),
+  lut[nb_steps] = static_cast<int16_t>(std::min<FloatT>(
+      std::max<FloatT>(TfLiteRound(LUTTransform<FloatT>(
+                                       transform, transform_params, input_max) *
+                                   output_scaling_inv),
                        table_min),
       table_max));
 }
 
+}  // namespace detail
+
+template <typename T>
+inline typename std::enable_if<std::is_same<T, uint8_t>::value ||
+                                   std::is_same<T, int8_t>::value,
+                               void>::type
+LUTPopulate(float input_scale, int32_t input_zero_point, float output_scale,
+            int32_t output_zero_point, float (*transform)(float), T* lut) {
+  detail::LUTPopulateInt8(input_scale, input_zero_point, output_scale,
+                          output_zero_point, transform, nullptr, lut);
+}
+
+template <typename T>
+inline typename std::enable_if<std::is_same<T, uint8_t>::value ||
+                                   std::is_same<T, int8_t>::value,
+                               void>::type
+LUTPopulate(float input_scale, int32_t input_zero_point, float output_scale,
+            int32_t output_zero_point, float (*transform)(float, const void*),
+            const void* transform_params, T* lut) {
+  detail::LUTPopulateInt8(input_scale, input_zero_point, output_scale,
+                          output_zero_point, transform, transform_params, lut);
+}
+
 template <typename T>
 inline typename std::enable_if<std::is_same<T, int16_t>::value, void>::type
 LUTPopulate(float input_scale, int32_t input_zero_point, float output_scale,
             int32_t output_zero_point, float (*transform)(float), T* lut) {
-  LUTPopulate<T, float>(input_scale, input_zero_point, output_scale,
-                        output_zero_point, transform, lut);
+  detail::LUTPopulateInt16<float>(input_scale, input_zero_point, output_scale,
+                                  output_zero_point, transform, nullptr, lut);
+}
+
+template <typename T>
+inline typename std::enable_if<std::is_same<T, int16_t>::value, void>::type
+LUTPopulate(float input_scale, int32_t input_zero_point, float output_scale,
+            int32_t output_zero_point, float (*transform)(float, const void*),
+            const void* transform_params, T* lut) {
+  detail::LUTPopulateInt16<float>(input_scale, input_zero_point, output_scale,
+                                  output_zero_point, transform,
+                                  transform_params, lut);
 }
 
-// Deprecated and will be removed in future, please use LUTPopulate instead
-template <typename FloatT, typename LutInT, typename LutOutT>
-inline void gen_lut(FloatT (*func)(FloatT), FloatT input_min, FloatT input_max,
-                    FloatT output_min, FloatT output_max, LutOutT* lut) {
-  static_assert(std::is_same<LutInT, LutOutT>::value,
-                "Input and output type of the LUT must be the same.");
-  static_assert(std::is_same<LutInT, int16_t>::value,
-                "Only int16_t type LUT are supported.");
-  static_assert(std::is_same<FloatT, float>::value,
-                "Only float type is supported for FloatT.");
-  using T = LutInT;
-
-  const auto zero_point = [](float min, float max, float scale) {
-    // Symmetric int16 LUT, we know the zero-point will not overflow an int32_t
-    // and zero-point from min will be the same as from max.
-    return static_cast<int32_t>(
-        static_cast<float>(std::numeric_limits<T>::min()) - min / scale);
-  };
-
-  const float scale = static_cast<float>(std::numeric_limits<T>::max() -
-                                         std::numeric_limits<T>::min());
-  const float input_scale = (input_max - input_min) / scale;
-  const FloatT output_scale = (output_max - output_min) / scale;
-  const int32_t input_zero_point =
-      zero_point(input_min, input_max, input_scale);
-  const int32_t output_zero_point =
-      zero_point(output_min, output_max, output_scale);
-
-  return LUTPopulate<T, float>(input_scale, input_zero_point, output_scale,
-                               output_zero_point, func, lut);
+// Deprecated, avoid usage and prefer the float version. Kept for
+// backward-compatiblity.
+template <typename T>
+inline typename std::enable_if<std::is_same<T, int16_t>::value, void>::type
+LUTPopulate(double input_scale, int32_t input_zero_point, double output_scale,
+            int32_t output_zero_point, double (*transform)(double), T* lut) {
+  detail::LUTPopulateInt16<double>(input_scale, input_zero_point, output_scale,
+                                   output_zero_point, transform, nullptr, lut);
+}
+
+// The size of the LUT depends on the type of input. For uint8 and int8 inputs a
+// simple 256 entries LUT is used. For int16 inputs the high 9 bits are used for
+// indexing and the 7 remaining bits are used for interpolation. We thus use a
+// 513-entries LUT for int16 cases, 512 for the 9-bit indexing and 1 extra entry
+// to interpolate the last value.
+template <typename T>
+constexpr int LUTSize() {
+  static_assert(std::is_same<T, uint8_t>::value ||
+                    std::is_same<T, int8_t>::value ||
+                    std::is_same<T, int16_t>::value,
+                "Only LUTs with uint8, int8 or int16 inputs are supported.");
+  // As per c++11: constexpr methods cannot have more than one return statement.
+  return (std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value)
+             ? 256
+             : 513;
 }
 
 // int16_t -> int16_t table lookup with interpolation

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

@@ -18,7 +18,6 @@ limitations under the License.
 #include <algorithm>
 
 #include "tensorflow/lite/kernels/internal/common.h"
-#include "tensorflow/lite/kernels/internal/portable_tensor_utils.h"
 
 namespace tflite {
 namespace reference_integer_ops {
@@ -134,20 +133,6 @@ inline void ConvPerChannel(
   }
 }
 
-inline void ConvPerChannelWithPackedInt4Weights(
-    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_input, int8_t* unpacked_filter_data,
-    const RuntimeShape& bias_shape, const int32_t* bias_data,
-    const RuntimeShape& output_shape, int8_t* output_data) {
-  TFLITE_DCHECK(unpacked_filter_data != nullptr);
-  tflite::tensor_utils::UnpackDenseInt4IntoInt8(
-      filter_input, filter_shape.FlatSize(), unpacked_filter_data);
-  ConvPerChannel(params, output_multiplier, output_shift, input_shape,
-                 input_data, filter_shape, unpacked_filter_data, bias_shape,
-                 bias_data, output_shape, output_data);
-}
 
 // Fixed-point per-channel-quantization convolution reference kernel.
 // 16-bit data and 8-bit filter

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

@@ -18,7 +18,6 @@ limitations under the License.
 #include <algorithm>
 
 #include "tensorflow/lite/kernels/internal/common.h"
-#include "tensorflow/lite/kernels/internal/portable_tensor_utils.h"
 
 namespace tflite {
 namespace reference_integer_ops {
@@ -122,21 +121,6 @@ inline void DepthwiseConvPerChannel(
   }
 }
 
-inline void DepthwiseConvPerChannelWithPackedInt4Weights(
-    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, int8_t* unpacked_filter_data,
-    const RuntimeShape& bias_shape, const int32_t* bias_data,
-    const RuntimeShape& output_shape, int8_t* output_data) {
-  TFLITE_DCHECK_NE(unpacked_filter_data, nullptr);
-  tflite::tensor_utils::UnpackDenseInt4IntoInt8(
-      filter_data, filter_shape.FlatSize(), unpacked_filter_data);
-  DepthwiseConvPerChannel(params, output_multiplier, output_shift, input_shape,
-                          input_data, filter_shape, unpacked_filter_data,
-                          bias_shape, bias_data, output_shape, output_data);
-}
-
 inline void DepthwiseConvPerChannel(
     const DepthwiseParams& params, const int32_t* output_multiplier,
     const int32_t* output_shift, const RuntimeShape& input_shape,